UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Modeling and analysis of TCP in satellite and mobile data networks interworking with the Internet Leung, Eva Y.F. 1998

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
831-ubc_1998-0330.pdf [ 5.34MB ]
Metadata
JSON: 831-1.0065315.json
JSON-LD: 831-1.0065315-ld.json
RDF/XML (Pretty): 831-1.0065315-rdf.xml
RDF/JSON: 831-1.0065315-rdf.json
Turtle: 831-1.0065315-turtle.txt
N-Triples: 831-1.0065315-rdf-ntriples.txt
Original Record: 831-1.0065315-source.json
Full Text
831-1.0065315-fulltext.txt
Citation
831-1.0065315.ris

Full Text

MODELING AND ANALYSIS OF TCP IN SATELLITE AND MOBILE DATA NETWORKS INTERWORKING WITH THE INTERNET by EVA Y. F. LEUNG B.A.Sc . (E .E. ) , University of Maryland, College Park, U . S . A . , 1993 A THESIS S U B M I T T E D IN P A R T I A L F U L F I L L M E N T O F T H E R E Q U I R E M E N T S F O R T H E D E G R E E O F M A S T E R O F A P P L I E D S C I E N C E in T H E F A C U L T Y O F G R A D U A T E STUDIES D E P A R T M E N T O F E L E C T R I C A L A N D C O M P U T E R E N G I N E E R I N G We accept this thesis as conforming to the required standard THE UNIVERSITY OF BRITISH COLUMBIA April, 1998 © Eva Y. F. Leung, 1998 F r o m : 04/Z7/6S 0fl:37 r'Ai 004 S22 9687 PHONE No. : 6 0 4 291 2 2 2 6 SPECIAL COLLECTIONS j A p r . 2 8 1 9 9 8 5 : 5 0 P M P 0 2 In presenting tttls thesis tn partial fulfilment of the requirement* for an advanced degree at the University of British Columbia, I agree that the Library shall make It freely available for reference and study. I further agree, that permission for extensive copying of this thesis for scholarly purposes may be | granted by the head of my department or by his or her representatives. Ii is understood that copying or publication of (his thesis for financial gain shall not be allowed without my written permission. The University of British Columbia Vancouver, Canada Date fri^f ^ /fZ£ DE-6 (2/88) Abstract The Transmission Control Protocol (TCP) is the most widely used transport protocol for the Internet because of its good performance over wireline networks. However, T C P performance is seriously degraded in many interconnected heterogeneous wireless and wired networks because their large delays and high packet loss rates violate many of the original T C P assumptions. This thesis presents an analysis and a comparison of the performance of different T C P implementations in networks with satellite links and mobile links interconnected with remote L A N s over the Internet, studies the effects of the critical network elements and the chosen T C P parameters on the T C P throughput performance, and obtains the optimal values for the chosen T C P parameters for the different T C P implementations. Compared with T C P Reno and T C P R F C , T C P Vegas not only has a better congestion avoidance mechanism to prevent network congestion and significantly reduce the probability of creating its own losses, but it also has a faster error recovery mechanism to quickly retransmit lost packets. Simulation results show that the satellite network elements and the T C P parameters have greater effect on the throughput of T C P Vegas than on the other implementations under most conditions for networks with satellite l inks . However, for networks with mobile l inks , the throughput performance of both T C P Vegas and T C P Reno is strongly affected by the mobile network elements and the T C P parameters. After jointly optimizing the key T C P parameters, the simulation results show that T C P Vegas has more than 100% better throughput than the other T C P implementations for networks with satellite links under a condition of high packet loss rates and long end-to-end delays. However, it just has an approximately 5% better throughput than T C P Reno for networks with mobile links under a condition of high packet loss rates and high mobility rates. ii Table of Contents Abstract ii List of Tables vi List of Figures vii Acknowledgment xi Chapter 1 Introduction 1 1.1 Backg round 1 1.2 Mot i va t i on 4 1.3 Prev ious W o r k 4 1.4 Object ive 8 1.5 Out l ine o f the Thesis 10 Chapter 2 Three TCP Implementations 11 2.1 The T C P Concept 11 2.2 Three T C P Implementations 16 2.2.1 Standard T C P 17 2.2.2 T C P Reno 19 2.2.3 T C P Vegas 20 2.2.4 Compar i son and Summary 22 Chapter 3 Heterogeneous Wireless and Wired Networking Environments 24 3.1 B roadband Satell ite Communica t ions 24 3.1.1 B roadband Satell ite M o d e l 25 3.2 M o b i l e Ce l lu l a r Communica t ions 27 3.2.1 M o b i l e Rad io M o d e l 28 iii iv 3.3 The Internet Environment 33 3.3.1 Internet Congestion Model 33 3.3.2 Internet delay model 36 Chapter 4 Analysis of TCP in a Satellite Environment 39 4.1 Satellite Network Interconnection v 39 4.2 Simulation Model 40 4.3 Simulation Assumptions and Parameters 41 4.4 Impact of Packet Loss Rate on T C P Throughput 42 4.5 Impact of End-to-End Delay on T C P Throughput 48 4.6 Impact of T C P Parameters on T C P Throughput 50 4.6.1 Effects of Maximum Transmit Window Size 51 4.6.2 Effects of Maximum A C K Delay 55 4.6.3 Effects of Maximum Segment Size 59 4.7 Optiminzation and Discussion 63 Chapter 5 Analysis of TCP in a Mobile Environment 73 5.1 Cellular Mobi le Communications 73 5.2 Simulation Model 74 5.3 Simulation Assumptions and Parameters 75 5.4 Impact of Mobi le Fading on T C P Throughput 76 5.5 Impact of Handoff on T C P Throughput 80 5.6 Impact of T C P Parameters on T C P Throughput 85 5.6.1 Effects of Maximum Transmit Window Size 85 5.6.2 Effects of Maximum A C K Delay 90 5.6.3 Effects of Maximum Segment Size 92 5.7 Optimization and Discussion 96 Chapter 6 Summary and Conclusions 104 6.1 Summary of Findings 104 6.2 Future Investigation 107 Bibliography 109 Appendix A. List of Abbreviations and Acronyms 113 Appendix B. Opnet Simulation Models 115 List of Tables Table 4.1 Fixed simulation parameters for the simulation model 42 Table 4.2 The optimal M T W S s for different T C P versions under different conditions 55 Table 4.3 The optimal maximum A C K delays for different T C P versions under different conditions 58 Table 4.4 The suggested optimal T C P paramenters for different T C P implementations 70 Table 4.5 Throughput in Kbytes/sec of T C P versions with the E T E delay of 0.35 sec 71 Table 4.6 Throughput in Kbytes/sec of different T C P versions at a P L R of 0.001 72 Table 5.1 The optimal transmit window sizes for both TCPs under different conditions 89 Table 5.2 The optimal M A D s for T C P Reno and T C P Vegas 92 Table 5.3 The Throughput of T C P Reno and T C P Vegas (Bytes/sec) 95 Table 5.4 The jointly optimal T C P parameters for T C P Vegas and T C P Reno 102 List of Figures Figure 2.1 The four layer of the TCP/IP protocol suite 11 Figure 2.2 T C P Data communication between a T C P source and a T C P destination 12 Figure 2.3 Encapsulation of T C P data in an IP datagram 14 Figure 2.4 The format of a T C P header 14 Figure 2.5 The T C P finite state simulation model 16 Figure 3.1 Broadband Satellite Model 25 Figure 3.2 Mobi le radio model 29 Figure 3.3 Occurrence of handoffs in a conventional cellular mobile system 30 Figure 3.4 The simulation Internet model 33 Figure 3.5 T C P application sources and Internet traffic model 35 Figure 3.6 Internet delay model 38 Figure 4.1 Satellite interconnected network topology 39 Figure 4.2 The simulation network model 41 Figure 4.3 Throughput of different T C P versions vs. the P L R in the satellite link 43 Figure 4.4 Congestion window response in different T C P implementations 45 Figure 4.5 The congestion window response in T C P Reno with different STPs 46 Figure 4.6 The congestion window response in T C P Vegas with different STPs 47 Figure 4.7 Throughput of different T C P versions vs. average E T E delay 49 Figure 4.8 Sequence number for data transfer over the satellite channel with a 0.001 P L R 50 Figure 4.9 Throughput of T C P R F C vs. M W T S for different P L R s 52 Figure 4.10 Throughput of T C P Reno vs. M W T S for different P L R s 53 vii Figure 4.11 Throughput of T C P Vegas vs. M W T S for different P L R s 53 Figure 4.12 Throughput of T C P Reno vs. M T W S for different E T E delays 53 Figure 4.13 Throughput of T C P Vegas vs. M T W S for different E T E delays 54 Figure 4.14 Throughput of T C P R F C vs. M A D for different P L R s 56 Figure 4.15 Throughput of T C P Reno vs. M A D for different P L R s 57 Figure 4.16 Throughput of T C P Vegas vs. M A D for different P L R s 57 Figure 4.17 Throughput of T C P Vegas vs. M A D for different E T E delays 57 Figure 4.18 Throughput of T C P Vegas vs. M A D for different E T E delays 58 Figure 4.19 Throughput of T C P R F C vs. the P L R in the satellite links for different M S S s ....62 Figure 4.20 Throughput of T C P Reno vs. the P L R in the satellite links for different MSSs.. . .62 Figure 4.21 Throughput of T C P Vegas vs. the P L R in the satellite links for different MSSs.. .63 Figure 4.22 Throughput of T C P Reno vs. P L R in the satellite links for different M S S s ( M T W S = 64 Kytes) 64 Figure 4.23 Throughput of T C P Reno vs. P L R in the satellite links for different M S S s ( M T W S = 80 Kbytes) 65 Figure 4.24 Throughput of T C P Reno vs. P L R in the satellite links for different M S S s ( M T W S = 128 Kbytes) 65 Figure 4.25 Throughput of T C P Reno with a segment size of 1458 bytes vs. the P L R in the satellite link 66 Figure 4.26 Throughput of T C P Vegas vs. P L R in the satellite links for different M S S s ( M T W S = 64 Kbytes) 67 Figure 4.27 Throughput of T C P Vegas vs. P L R in satellite links for different M S S s ( M T W S = 80 Kbytes) 68 Figure 4.28 Throughput of T C P Vegas vs .PLR in the satellite links for different M S S s ( M T W S = 128 Kbytes) .68 Figure 4.29 Throughput of T C P Reno with a segment size of 1458 bytes vs. the P L R in the satellite links 69 viii Figure 4.30 Throughput of different T C P implementations vs. the P L R in the satellite channel 71 Figure 4.31 Throughput of different T C P implementations vs. the average E T E delay a taO.OOlPLR 72 Figure 5.1 Interconnection between a mobile host and Internet fixed hosts 73 Figure 5.2 The mobile data network simulation model 75 Figure 5.3 Average P L R in the mobile radio channel vs. mobile speed 77 Figure 5.4 T C P Throughput vs. mobile speed for different P L R s in the bad state 78 Figure 5.5 T C P throughput vs. P L R in the bad state for different mobile speeds 79 Figure 5.6 The congestion window response in T C P for different mobile speeds 79 Figure 5.7 T C P throughput vs 1/a for different 1/b values at a mobile speed of 45 K m / h 81 Figure 5.8 T C P throughput vs 1/a for different 1/b values at a mobile speed of 75 K m / h 81 Figure 5.9 T C P throughput vs 1/a for different 1/b values at a mobile speed of 120 K m / h ...82 Figure 5.10 T C P throughput vs. 1/a for different 1/b values at mobile speed of 75 K m / h 83 Figure 5.11 Sequence numbers for T C P segments transferred to mobile host over the mobile channel 84 Figure 5.12 Congestion window response in T C P over mobile channel at a mobile speed of 75 Km/h 84 Figure 5.13 T C P throughput vs. M T W S for different mobile speeds 86 Figure 5.14 T C P throughput vs M T W S for different 1/a values 87 Figure 5.15 T C P throughput vs. M T W S for different 1 lb values 88 Figure 5.16 T C P throughput vs. M T W S for 1/a = 140 sec 89 Figure 5.17 T C P throughput vs M A D at different mobile speeds 90 Figure 5.18 Throughput of T C P Reno vs. M A D for different M T W S s 91 Figure 5.19 Throughput of T C P Vegas vs. M A D for different M T W S s 91 ix Figure 5.20 T C P throughput vs the M S S for different mobile speeds 94 Figure 5.21 The throughput of T C P Reno vs 1/a for different M S S s 94 Figure 5.22 The throughput of T C P Reno vs 1/a for different M S S s 94 Figure 5.23 Throughput of T C P Reno with a 256 Kbyte M S S vs. 1/a for different M A D s 97 Figure 5.24 Throughput of T C P Reno with a 512 byte M S S vs. 1/a for different M A D s 97 Figure 5.25 Throughput of T C P Reno with a 896 byte M S S vs. 1/a for different M A D s 98 Figure 5.26 Throughput of T C P Reno with a 1024 byte M S S vs. 1/a for different M A D s 98 Figure 5.27 Throughput of T C P Vegas with a 256 byte M S S vs. 1/a for different M A D s . . . . 100 Figure 5.28 Throughput of T C P Vegas with a 512 byte M S S vs. 1/a for different M A D s 100 Figure 5.29 Throughput of T C P Vegas with a 896 byte M S S vs. 1/a for different M A D s 101 Figure 5.30 Throughput of T C P Vegas with a 1024 byte M S S vs. 1/a for different M A D s . . . 101 Figure 5.31 T C P throughput vs. 1/a : 102 Figure 6.1 The throughput comparison of T C P Vegas and a proposed T C P vs. the P L R in the satellite links 108 Figure 6.2 The throughput comparison of T C P Vegas and a proposed T C P vs. the P L R in the bad state within the normal state for networks with mobile links 108 Acknowledgment This thesis is dedicated to my parents, Fong Leung and Mee Wah Chu, and my brothers and sisters for continuous support and encouragement for me to pursue my goal. In particular, I would like to express my sincere gratitude to my research supervisor, Dr. Victor C . M . Leung, for his guidance and constant patience throughout my graduate studies at the University of Bri t ish Columbia. Lastly, I would offer my appreciation to my close friends and my colleagues for their continuous support. This work was funded by the Canadian Institute of Telecommunications Research (CITR) and Motorola, Inc. through a Research Assistantship provided by Dr. Victor Leung. I would also like to thank M i l 3, Inc. for providing the state-of-art simulation software, O P N E T , and technical support throughout this work. xii Chapter 1 Introduction Over the last decade Internet communication has been growing rapidly, providing high-speed network connectivity to a large number of heterogeneous computer applications. However, the existing terrestrial wide area network is wired, and therefore l imited both in its wide area coverage and by its inability to link with mobile hosts. To transcend these limitations, wireless communications technologies have been developed for interconnecting stationary or mobile remote hosts. Broadband satellite communications can provide wide area coverage, typically to stationary hosts. Land mobile cellular communications operate over smaller areas, but can l ink stationary (e.g. Internet) to mobile hosts. Analyses of Internet applications [1][27] indicate that the packet length, terminal activity, and data rate of user terminals may keep changing by many orders of magnitude, thus causing the Internet traffic to change dynamically over a wide range. Large file transfer, multimedia communication, and image conferencing are expected to be the major future network traffic. These applications wi l l require short response time, high throughput performance, and very flexible interconnection architectures. Therefore, reliable transport protocol performance and flow control analyses of communi-cations end-to-end across both wireless and wired networks are critically needed to ensure the desired quality of service in seamless end-to-end delivery of the data traffic. Because Transmis-sion Control Protocol (TCP) works very well in traditional wireline networks made up of wired links and stationary hosts, it has been considered for use on heterogeneous wireless and wired ( H W W ) networking environments. 1.1 Background T C P is currently the most widely used transport protocol for a large spectra of topologies 1 Chapter 1 Introduction 2 in wide area networks or in the Internet. T C P [20] [28] was originally designed for the slow-speed, unreliable, datagram-based A R P A N E T where packets were often lost, duplicated or delivered out of sequence. Because computing and networking technologies have been growing and the characteristics of the Internet traffic have been significantly changing, many implementa-tions and options for T C P have been introduced to enhance the service of the Internet Protocol (IP). The different T C P implementations are presented in detail in chapter 2. T C P employs a window scheme to control packet flow from a sender to a receiver. It also uses a positive acknowledgment with a retransmission scheme for error recovery. B y adapting to end-to-end delay and packet loss rate, the error recovery and congestion control mechanisms in T C P have been performing very well in low bit error rate wireline networks such as the Internet. In these networks packet losses are mainly due to network congestion. However, communications over satellite or mobile radio networks are quite different from those over traditional wireline networks [16]. These wireless links do not have the high bandwidth of wired links, have higher latencies, and have much higher error rates. Their packet losses, in contrast to the wi red networks, are mainly caused by outages due to connection interruptions. In broadband satellite channels, large data segments can be lost when the channel changes suddenly due to a rapid change in the weather conditions. In the mobile environment, the packet losses are mainly caused by wireless links which are temporarily aborted due to deep fades or due to hard handoffs when the mobile crosses a cell boundary. Because of these unexpected interruptions, the error recovery and f low con t ro l mechanisms in T C P cannot work well . T C P always misinterprets an unexpected increase in delay as packet loss due to network congestion [2] [3] [4] [5] [6] [7] [10]. A s a result, T C P first drops the transmission window size to reduce the amount of data in transit through the network. Second, it activates its slow-start algorithm to control the traffic rate at which the traffic grows back to its Chapter 1 Introduction 3 previous level. Third, it resets its retransmission timer to a back off interval that doubles with each consecutive time-out. If a packet is dropped due to an overloading of limited resources in the network, conges-tion control has to be applied immediately in order to restrict the traffic flow and give the network a chance to clear up. However, i f a connection loses a packet due to the unreliable wireless links, which means the packet was corrupted by errors or by a temporary communication pause due to a handoff process, no congestion control is necessary. The sender should not reduce the data flow rate by performing the slow-start mechanism to drastically decrease the congestion window size, thus wasting the available bandwidth and further degrading the system performance. The problem is worsened in severely fading channels [3]. The slow start mechanism w i l l be highly active for a long period of time, thus increasing the frequency of end-to-end communications pauses. Because of the exponential back-off policy, the communication pauses at the sender w i l l be unnecessarily long, especially in high mobility environments [7][11]. Although the slow-start mechanism works well for avoiding network congestion when it is used over terrestrial wireline network (low packet loss rate and small delay), it is very inefficient in high loss rate and highly interactive delay environments. The long propagation delay in broadband satellite networks is another major problem degrading T C P throughput. Wi th a long propagation delay between a source and a destination, the frequency of sending packets into the network is low because the sender must wait for a long round-trip-delay to receive an A C K before it can send or retransmit packets. If it is using the slow-start mechanism for congestion prevention, T C P further l imits the size of the transmit window for every packet loss detected by time-out. Therefore, the utilization of a long propaga-tion delay satellite channel is degraded by the limited T C P window size in combination with the T C P maximum segment size. Chapter 1 Introduction 1.2 Motivation 4 Recently there has been significant increase in activity in the area of wireless communica-tions with both stationary and non-stationary hosts interworked with the Internet. T C P is the most widely used reliable transport protocol and has recently been considered for supporting hosts in heterogeneous networks. However , T C P performance is seriously degraded in wireless communications environments because networks with wireless links and mobile hosts seriously violate many of the original T C P assumptions. In order to ensure the desired quality of service for the data traffic and to exploit the inherent strengths of the heterogeneous networks, it is necessary to perform a detailed perfor-mance analysis of T C P in the heterogeneous wireless and wired ( H W W ) data communications networks under various conditions. T C P has gone through several refinements to improve perfor-mance in the highly dynamic Internet traffic environments. The core mechanisms such as flow control and error recovery in different T C P implementations have been modified. Therefore, the performance analysis of different T C P implementations for the networks with satellite links and mobile l inks interconnected with remote L A N s over the Internet is needed. Because a lot of mutually dependent T C P parameters strongly affect the T C P performance, optimization of T C P parameters is also needed for different T C P implementations to yield better throughput perfor-mance. 1.3 Previous Work Many performance analyses have been done for T C P / I P in an interconnection satellite network environment and in a mobile computing environment. A number of studies have been reported in the literature dealing with various aspects required to support hosts in the satellite network environment and mobile computing environments. Several design alternatives have been Chapter 1 Introduction 5 proposed to improve the poor end-to-end T C P performance over a network with wireless links. The authors in [5] presented the performance evaluations in the areas of satellite l ink delays and bit error rates for the T C P / I P protocol stack and the X T P implementation. In this paper, the T C P implementation based its error recovery strategy on the a Go-back-N protocol; while the X T P [12] implementation based its error recovery strategy on the selective repeat protocol with a sliding window flow control. From the experimental measurement, the authors' conclusions are: 1) The window size has a stronger influence on T C P throughput than on X T P throughput. Large window sizes provide a better channel uti l ization on high delay l inks at a relatively low bit error rate. 2) A retransmission ambiguity problem of the T C P implementation makes it more sensitive to a larger delay condition than the X T P implementation. 3) X T P has better performance than T C P due to the use of a selective repeat algorithm of retransmissions for corrupted or lost packets even under a high bit error rate condition. The slow start mechanism of T C P is the main reason for degradation of its performance in H W W networks. In [3], the authors analyzed the throughput performance of different T C P error control strategies, which are the N A K option, the selective repeat option, and the Shutter X O R selective repeat strategy, while switching off the T C P congestion control scheme for L A N interconnection through a satellite channel. In the simulation the authors found that the throughput was increased by explicit error detection mechanisms because of the reduction of unnecessary retransmissions. The selective repeat option has a higher throughput than the N A K option in the case of high bit error probabilities over a bandwidth-delay dominated network, and the Shutter-XOR selective repeat strategy has slightly better throughput than the selective repeat option at a high bit error rate. A t a bit error rate of 5x10~ 6, the optimal value of the maximum window size for T C P with selective repeat error control strategies is about 64 Kilobytes under the simulation conditions. The throughput efficiency cannot be increased by further increasing this value. Chapter 1 Introduction 6 The authors of [2] [21] found that the transport protocol are very sensitive to probability of bit error and the altitude of the satellites. Bigger window sizes can improve the end-to-end throughput performance at a relatively low bit error rate. In [4], some key observations were made about the file transfer in a L A N interconnection through a satellite channel. From the simulation results, the authors found that the flow control window in the T C P / I P is the bottleneck of the system i f the bridge is implemented for the use of a single pair of workstations. However, the bottleneck wi l l move to the bridge due to its queueing delay, especially at a smaller window size, if the bridge is shared by multiple workstation pairs. In [7], the effects of mobil i ty on performance of the TCP-Tahoe implementation was studied by measuring T C P behavior in a wireless networking testbed. A T C P connection was initiated between a mobile host connected to a 2-Mbits/sec W a v e L A N local area network and a stationary host connected to a 10-Mbits/sec Ethernet local area network. The motion across cell boundaries was simulated with the mobile host switching cells every 8 beaconing periods. The authors found that cell crossings cause increased delays and packet losses. T C P interprets this delay as an indication of network congestion, thus activating the slow-start algorithm to restrict the traffic flow. Because of this misinterpretation, the throughput performance of T C P drops significantly with increased mobile host motion. During handoff pauses, T C P transmits no new data and transport-level communication comes to a halt; as a result, the throughput performance of T C P for a 1-second rendezvous delay is degraded to approximately one-third of the T C P throughput under the no handoff condition. The authors of [7] also investigated the effect of having a fast retransmit approach described in [10] in the T C P implementation, and showed that the mechanism could reduce the pauses in communication from 2.8 to 1.2 seconds for a 1-second rendezvous delay. However, this approach, which is dependent on network layer information, can only deal with handoffs but not with the error characteristics of the wireless links. Chapter I Introduction 7 In [6], a simulation study of the impact of mobili ty on performance of the TCP-Tahoe implementation is presented. A queueing network model with the characteristics of Internet delay is developed. F r o m the simulation results, the authors found that the packet losses during handoffs can significantly drop the throughput efficiency especially in highly mobile environ-ments and high bandwidth connections. Large window sizes such as 64Kbyte, further reduce the mean throughput by approximately 10 percent even in a host-limited network. The authors concluded that the loss of throughput is mainly due to the fact that the T C P implementation misinterprets packet loss due to mobility as an indication of congestion in one or more network resources. They also proposed a simple T C P protocol modification which uses the network layer information of any ongoing mobility to extend the slow-start phase in the recovery process. The authors showed that the modified T C P protocol significantly improves the throughput for host limited T C P connections in the case of a typical 0.15 seconds handoff interruption for every 120 second. There is a 20 percent increase in the case of large window sizes. Another performance analysis of two T C P implementations in a wireless environment can be found in [8]. In this paper, the authors studied the protocol I, which is similar to TCP-Tahoe except it uses a fine-grained timer, and protocol II, which is a variant of T C P Vegas [26], in a simulation model with the similar characteristics similar to the measurement testbed in [7]. In the simulation, periodic and exponential intermobility times with a high bit error rate were consid-ered. From the simulation results, the authors showed that both protocols performed, in general, badly when the wireless channel was not available 50% of the time, but protocol II was more robust than Protocol I when the wireless channel was available more than 50% of the time. The authors also concluded that the larger segment size gave slightly better throughput when there were no random errors but significant improvement in throughput at 5% random packet error rate. The I - T C P approach described in [13] involves splitting the T C P connection into two Chapter I Introduction 8 separate connections at the base station, one between the mobile host and the base station and the other between the base station and the mobile host. The advantage of this approach is to separate the flow and congestion control of the wireless links from that of the Internet; however, there are also many disadvantages of this approach which are the semantics problem, the special socket system and the software overhead. Another approach is to have its own retransmission protocol implemented at the data link level; however, studies have shown that this implementation leads to further degraded performance in the wireless environment [14]. 1.4 Objective Most of the earlier studies on the performance of T C P used simplified T C P versions over either wireless or wired links modeled by a tandem network of queues with a single traffic type. This thesis focuses on the performance of three different T C P implementations in wireless networking environments interworked with the Internet. In order to have a better T C P throughput performance, optimization of T C P parameters for the heterogeneous wireless and wired network under different conditions is considered. In this thesis, two cases are considered: two remote L A N ' s l inked by a satellite network interworked with the Internet; and a mobile host intercon-nected with an Internet fixed host by mobile radio link in a Cellular Digital Packet Data ( C D P D ) environment [33]. It is important to abstract clearly the characteristics of the Internet and both wireless networking environments in order to carry out a detailed analysis of their impact on different T C P implementations. A lot of mutually dependent parameters have a great influence on the performance of T C P in the networking environments; therefore, the ultimate choice of these parameters and their value wi l l seriously affect the T C P performance for bulk data transfer. Chapter 1 Introduction 9 The objectives of the thesis are: • to develop a queueing network model with the empirical characteristics of the Internet, the broadband satellite link, and mobile radio links; • to analyze the effect of some key T C P parameters of the different T C P implementa-tions in the heterogeneous networking environments with different conditions; • to find the optimal values of the key parameters for the T C P implementations in the re-spective networking environments; and • to provide suggestions in modifications of T C P for better throughput performance in both H W W networking environments. This work differs from previous investigations in the following important ways: • Three full versions of T C P (the standard version of TCP, T C P Reno, and T C P Vegas) are investigated and analyzed for networks with satellite links and mobile links inter-working with the Internet. • The Internet model is based on the empirical characteristics of wide area TCP/ IP con-versations, data packet traffic, and end-to-end behavior. • A two-state Markov model is used for modeling a fading satellite channel. • In the mobile computing environment, a three-state continuous time Markov model is used for modeling a fast fading mobile radio channel with frequent handoff interrup-tions. • The jointly optimal values for the key T C P parameters are presented in both networks with satellite links and land mobile links. Chapter 1 Introduction 1.5 Outline of the Thesis 10 In chapter 2, an overview of different T C P implementations is presented and discussed. Chapter 3 provides an overview of the wireless data communications systems, which are a broadband satellite communications network and a land mobile radio computing network, and the generic Internet environment with the relevant network models. Chapter 4 presents the simulation results and the performance analysis of different T C P implementations in a satellite network environment interworked with the Internet. The optimization of the chosen T C P parameters for the T C P implementations in this broadband satellite network is also given here. Simulation results and performance analysis for mobile computing networks with the various T C P implementations are given in Chapter 5. The optimization of the chosen T C P parameters in this environment is also presented here. Final ly , Chapter 6 summarizes al l the findings for both types of H W W networks, and provides suggestions for future work on improving T C P throughput performance in the networks with wireless links interconnected with the Internet. Chapter 2 Three TCP Implementations In this chapter, an overview of the generic T C P concept is presented. Three different T C P implementations including the standard version of TCP, T C P Reno, and T C P Vegas are studied and discussed in detail. T C P / I P protocol suite is a networking protocol suite and the combination of different protocols at various layers. Figure 2.1 shows the 4-layer network system for TCP/ IP . This protocol suite allows different kinds of computers, running on different operation systems, to communicate with each other over the worldwide Internet. Each layer of the T C P / I P protocol suite has different responsibilities. The link layer mainly handle all hardware details of physically interfacing with different types of media that is being used. The network layer handles packet routing over the worldwide Internet. The transport layer provides a flow of data between two hosts for the application layer above. The application layer handles the details of the particular application and user processes. 2.1 The TCP Concept Application FTP, N N T P , Telnet, etc. Transport TCP, U D P Network IP, I C M P , I G M P Link device driver and interface card Figure 2 .1 The four layer of the TCP/IP protocol suite. 1 1 Chapter 2 Three TCP Implementations 12 Internet Protocol (IP) is a commonly use network protocol to route packets as a datagram by its routing mechanism. Since IP only provides a connectionless packet delivery service to route each packet separately, it has no guarantee of reliability or in-order delivery for each packet, which means that packet can be lost or destroyed by the media such as when network hardware fails, or packet can be substantially delayed by the dynamical network routing. Transmission Control Protocol (TCP) , which is a commonly used transport protocol, has a responsibility for providing a reliable flow of data between two hosts. Therefore, the application layer can ignore all the details of the data rel iabil i ty issue. Figure 2.2 shows the T C P data communications between two remote T C P hosts and all the protocols involved. T C P accepts files of any lengths from applications and breaks them up into smaller segments. Each T C P data segment is encapsu-lated in an IP datagram and sent from a source to a destination through the lower layers of the L A N ' s . Application Application Application Application s^ocket^  s^ocket^  • T C P • Virtual Connection s^ocket^  s^ocket^  • T C P • i — — — — — — — — — — — — — — — Internet (IP) Internet (IP) 1 1 Network Interface Network Interface Source Destination Figure 2.2 TCP Data communication between a TCP source and a TCP destination T C P is a connection-oriented and end-to-end reliable transport protocol to support multi-network applications. In term of connection-orientation, two remote T C P applications must establish a T C P connection with each other before they can exchange data. Us ing a three-way Chapter 2 Three TCP Implementations 13 handshake mechanism with clock-based sequence numbers not only can avoid erroneous initial-ization of connections but also secure a unique end-to-end connection between two hosts. Since the specified sequence numbers are sent and acknowledged during the handshake, the handshake mechanism further guarantees that both sides are ready to transfer data and agree on the initial specified sequence numbers. When the connection is ready to close, T C P uses a modified three-way handshake to close the connection. T C P w i l l close the connection in one direction when there is no more data to send, and then close the other half of the connection when it receives an acknowledgment that no more data is available. In term of data reliability, T C P provides several main operations including basic data transfer, error recovery, and flow control. TCP, in general, decides when to block and send data at its own convenience to maintain a continuous stream of bytes in each direction between two end hosts. A stream of 8-bit bytes with no record markers inserted by T C P is exchanged across the T C P connection between two T C P applications. In this byte stream service, one end puts a stream of bytes into TCP, and the same and identical stream of bytes appears at the other end. Figure 2.3 and figure 2.4 show the encapsulation of T C P data in the IP datagram and the format of the T C P header, respectively. T C P segments are used to establish connections as well as to carry data and acknowledg-ments. Each T C P segment is divided into two parts, the T C P header and the actual data. The T C P header only carries the expected identification and the control information. The "Source Port" and "Destination Port" fields identify T C P port numbers of the application programs at both ends of the connection. The "Sequence Number" field identifies the position of the segment in the sender's byte stream. The "Acknowledgment Number" field identifies the number of bytes that the source expects to receive next. The " H L E N " field contains an integer that specifies the length of the segment header measured in 32-bit multiples. The 6-bit "Reserve" filed is reserved for Chapter 2 Three TCP Implementations 14 future use. The 6-bit "Code Bi ts" is used by T C P software to determine the purpose and contents of the segment. The "Window Size" field is to specify the current receive buffer size on the other end. The "Checksum" field is used for error detection. The "Urgent Pointer" is used to specify the position in the window where urgent data ends. The "Option" field is used to negotiate with the T C P software on the other end of the connection. ^ IP datagram • ~< TCP Segment • IP T C P Header Header T C P Data Figure 2.3 Encapsulation of TCP data in an IP datagram Source Port Destination Port Sequence Number Acknowledgment Number H L E N Reserved Code Bits Window Size Checksum Urgent Point Options (if Any) 20 bytes Figure 2.4 The format of a TCP header In order to transport data reliably, a T C P sender uses a timer to maintain a time-out for a sending segment. A T C P receiver sends an A C K after it receives a correct data segment from the sender. If a corresponding acknowledgment from the receiver is not received in time, the segment is retransmitted. T C P also maintains a checksum on its header and data to check i f there is any modification of the data in transit. It discards the damaged data segment and does not acknowl-edge receiving it. Moreover, it resequences the out of order data segments i f necessary and passes Chapter 2 Three TCP Implementations 15 them in a correct order to the application. B y using the positive acknowledgment, It can recover from data transmission errors such as damaged, lost, or duplicated packets over the unreliable Internet. To prevent congesting a network and creating further losses, T C P provides efficient flow control. T C P senders uses a maximum allowable window size indicated in a received T C P header to control the amount of data that can be sent into the network. Congestion is a condition of severe delay caused by an overload of datagrams at one or more switching points. Therefore, a combined slow-start with congestion avoidance algorithm [17], referred as the slow-start mechanism, is implemented in T C P to further control the data traffic flow. The overview of the congestion avoidance and slow start algorithms are presented in section 2.2.1. The operation of T C P can be explained with a finite state model. Figure 2.5 shows the finite state model of T C P , with circles representing states and arrows representing transitions between them. The label on each transition shows what T C P receives to cause the transition and what it sends in response. Applications must issue either a passive open command, waiting for a connection from another T C P host, or an active open command, initiating a connection to the destination. A n active open forces a transition from the init state to the S Y N _ S E N T state. After the three-hand shake mechanism ( S Y N . S E N T and S Y N _ R C V ) , T C P moves to the E S T A B state and begins data transfer. A l l operations for data reliability such as the flow control and the error recovery perform in the E S T A B . To avoid having segments from a previous connection interfere with a current one, T C P moves to the T I M E _ W A I T state after closing a connection. It remains in these states ( F I N W A I T I and II) for twice the maximum segment lifetime before deleting its record of connection. To handle cases where the last acknowledgment was lost, it acknowledges valid segments and restarts the timer. Because the timer allows T C P to distinguish old connec-tions from new ones, it prevents T C P from responding with a reset i f the other end retransmits a Chapter 2 Three TCP Implementations F I N request. 16 2.2 Three TCP Implementations T C P described in [28] is referred as the standard version of T C P since it has been implemented in many commercial products over many years. Because the computer networking environments and the characteristics of the Internet traffic have been changing substantially, many options and modified T C P implementations have been proposed. However, the flow control and error recovery mechanisms are only slightly modified. Three versions of T C P studied here are the standard version of T C P (TCP R F C ) , T C P Reno, and T C P Vegas. Chapter 2 Three TCP Implementations 2.2.1 Standard TCP 17 T C P R F C was o r ig ina l ly designed for a s low-speed, unre l iable , datagram-based A R P A N E T where packets were often lost, duplicated or delivered out of sequence. In T C P R F C [28], the retransmission time out (RTO) algorithm is based on using the round trip time of each segment, a smoothing factor, and a delay variance factor to update a current round trip time (RTT) and to estimate a R T O for unacknowledged segments. R T T is measured by the elapsed time between sending a data byte with a particular sequence number and receiving an acknowledgment that covers that sequence number. A Smoothed Round Trip Time (SRTT) is computed as: SRTT = ( a * SRTT) + ((1-a) * RTT) and used to update R T O calculated as: R T O = min [ U B O U N D , max [ L B O U N D , ((3 *SRTT)]], where U B O U N D is an upper bound on the time-out and L B O U N D is a lower bound on the time-out, a is a smoothing factor between 0.8 and 0.9, and (3 is a delay variance factor between 1.3 and 2.0. G o - b a c k - N A R Q scheme is used; therefore, i f a segment is lost in transit, T C P R F C retransmits all the outstanding segments starting from the sequence number of the lost segment even through some of these packets may have been received correctly. In networks with high packet loss rates and long end-to-end delays, this mechanism significantly reduces T C P through-put because it may create more losses by further congesting the network. Lost segments in the T C P congestion avoidance algorithm are used as a signal of network congestion. Congestion avoidance and slow start are independent algorithms with different objectives. They both require two variables, which are a congestion window, called cwnd, and a slow start threshold size, called ssthresh, to maintain the connectivity and the traffic flow between two end hosts. The congestion window is a second l imit window employed by the slow start mechanism at the T C P sender to avoid network congestion. Whenever starting traffic on a new Chapter 2 Three TCP Implementations 18 connection, cwnd is initialized to one segment and ssthresh is typically initialized to 64 Kbytes. Cwnd is increased by one segment for every non-duplicated acknowledgment received. This provides an exponential increase in traffic flow. On a non-congested connection, the congestion window size in the sender is the same as the receiver window. The sender can transmit up to the min imum of the congestion window and the advertised window. Therefore, the congestion window is a type of flow control presumed by the sender while the advertised window is a type of flow control presumed by the receiver. To estimate congestion window size when losses occur (indicated by a time-out or the reception of duplicated A C K s ) , T C P assumes most losses come from congestions and reduces its congestion window size. It saves one-half of its current window size in sshtresh, and cwnd is set to one segment only i f congestion is indicated by a time-out, i.e., slow start. For those segments that remain in its allowed window, the R T O is backedoff exponentially, according to the exponen-tial backoff algorithm. When new packet is acknowledged by the destination, increase in cwnd depends on whether T C P is performing slow start or congestion avoidances. If cwnd < ssthresh, T C P is performing slow start; otherwise, T C P is performing congestion avoidance. Slow start has cwnd beginning at one segment and exponentially increasing until it is halfway (i.e, equal to the new ssthresh) to where it was when congestion occurred, and then congestion avoidance takes over the flow control. Slow start exponentially increases cwnd by the number of A C K s received in a round-trip time after it is set to be one segment when time-out occurs. The increment of cwnd in congestion avoidance for every received A C K is estimated as equation (1), cal led linear increase, where segsize is the maximum T C P segment size. . 2 , , , segsize , segsize m cwnd = cwnd + — — + — (1; cwnd 8 Chapter 2 Three TCP Implementations 2.2.2 TCP Reno 19 T C P Reno is the most commonly used T C P implementation in today's systems. The R T O mechanism in T C P Reno [27] is refined by an R T T gain, an R T T coefficient deviation, and an estimator of an average of an R T T [17]. The measurement of an R T T in T C P Reno is the same as that in T C P R F C . A measured R T T for a particular sequence number is denoted by M . A n updated R T O that is applied to M for the next transmitting segments is estimated as Err = M-A A - A + gErr D = D + h(\Err\-D) A = A + rttcoef x D where A, D, g, h, rttcoef, and Err represent, respectively, a smoothed RTT, a smoothed mean deviation, an average gain (set to 0.125), a deviation gain (set to 0.25), an R T T coefficient deviation gain (set to 4), and the difference between a measured value just obtained and a current R T T estimator. Therefore, a larger deviation gain for a deviation can make an R T O go up faster when an R T T changes. Karn's algorithm [11] is implemented to improve the T C P retransmission ambiguity problems. This algorithm specifies that when a time-out and a retransmission occur, the R T T estimators cannot be updated when the acknowledgment for the retransmitted data finally arrives. The basic error detection mechanisms are based on a specified T C P segment timer maintained at a transmitter and the third duplicated A C K . A transmitter's system clock is recorded only when a selected segments are sent, and an R T O is calculated and updated only when an corresponding A C K s of these segments are received. The original retransmit time-out Chapter 2 Three TCP Implementations 20 algorithm is used, but checking for time-out occurs only when a coarse grain timer with a resolu-tion of 500 ms has expired. A fast retransmit and fast recovery ( F R F R ) algorithm [17] [18] is implemented to improve the original retransmission mechanism. This algorithm uses the third duplicated A C K as an indication of the need to retransmit a lost segment. When a transmitter receives three duplicated A C K in a row from a remote receiver, F R F R procedure immediately retransmits the earliest unacknowledged packet. This algorithm sets ssthresh to one-half of the current cwnd, retransmits the lost segment, and sets cwnd to the sum of the current ssthresh and three times of the maximum segment size. Cwnd is increased by one segment size for every duplicated A C K arrival; however, cwnd is only set to ssthresh when the next (non-duplicated) A C K arrives after an retransmission. T C P Reno still uses the combined slow start and congestion avoidance algorithm described in section 2.2.1 to control the data traffic flow into the network. A selective repeat A R Q scheme is used; therefore, the request of retransmissions from the sender apply only for those packets that are not correctly received. A buffer is implemented in the T C P receiver to save and sort all the out-of-order packets. In the receiver, T C P saves the out-of-order received packets in the receiver buffer and responds with an A C K of the highest sequence number successfully received, plus one segment size. If the next received segments are also out of order, the data is saved and duplicated A C K s are generated. When the missing data segment arrives, the receiver now has in-sequence data segments starting from the missing data segment and passes these segments in order to the user process. 2.2.3 TCP Vegas T C P Vegas described in [26] is a newly proposed T C P implementation which is based on T C P Reno. There are several modifications in retransmission, congestion avoidance, and slow-start mechanisms in T C P Vegas in order to improve throughput performance and decrease losses. Chapter 2 Three TCP Implementations 21 In the new retransmission mechanism, a finer-grained timer is used to calculate a more accurate R T T estimation. Its system clock is recorded each time a segment is sent, and the R T T is updated using the A C K arrival clock and its domestic clock recorded for the respective segment. Thus, T C P Vegas extends the F R F R algorithm implemented in T C P Reno. B y using this more accurate R T T estimation as a time-out value for a specified packet, T C P Vegas decides to retransmit segments only under two situations. First, when a duplicate A C K is received, it checks to see i f the difference between the current time and the timestamp recorded for the relevant segment is greater than the time-out value for the relevant segment. If it is, then T C P Vegas retransmits the segment without having to wait for the arrival of the third duplicated A C K . Second, when a non-duplicated A C K is received after a retransmission, T C P Vegas again checks to see i f the time interval since the segment was sent is larger than the time-out value. If it is, then T C P Vegas retransmits the segment. Us ing these two retransmission approaches, T C P Vegas can catch any other segments that may have been lost previous to the retransmission without having to wait for a duplicate A C K . It only decreases the congestion window i f the retransmitted segment was previously sent after the last decrease. Coarse-grain time-out as in T C P Reno is still employed in case the above mechanisms fail to recognize a lost segment. The same selective repeat A R Q in Reno is used to retransmit the lost segments only. In T C P Vegas, the congestion avoidance are based not only on dropped segments but also on changes in an estimated amount of extra data in the network. T C P Vegas defines the BaseRTT as the R T T of a data packet when the connection is not congested. It is the min imum of a l l measured round trip times. Under a good condition, the expected throughput should be equal to the size of the current congestion window (the number of bytes in transit) d iv ided by the BaseRTT. The actual throughput is the number of bytes transmitted between the time that a specified segment is sent and its acknowledgment is received divided by the recorded sending Chapter 2 Three TCP Implementations time for the distinguished segment. 22 T C P Vegas keeps the congestion window size unchanged when the difference between the expected throughput and the actual throughput is in the range of an upper and a lower threshold. Otherwise, it modifies the congestion window size by a linear increase or decrease as in T C P Reno. These thresholds are the multiple of a maximum segment size divided by the BaseRTT. The lower threshold allows the connection to utilize at least a certain number of buffers at a bottle-neck router in the system, and the upper threshold ensures the connection uses no more than a certain number of buffers in the network. If the actual throughput is greater than the expected throughput, the BaseRTT is changed to the latest sampled RTT. When the difference between the expect throughput and the actual throughput, referred as Diff, is less than the lower threshold, T C P Vegas increases the congestion window linearly during the next R T T (see equation 1). When D i f f is greater than the upper threshold, T C P Vegas decreases the congestion window linearly during the next RTT. In order to detect and avoid congestion during slow-start, T C P Vegas allows exponential growth in the congestion window only every other RTT. In between, the congestion window stays fixed; therefore, a valid comparison of the expected and actual throughput can be calculated. 2.2.4 Comparison and Summary T C P R F C and T C P Reno have a similar congestion control mechanism except for the F R F R mechanisms implemented in T C P Reno. They have no mechanism to detect the onset of network congestion before packet losses occur, and thus they create their own losses due to buffer overflow by continually increasing window size. It is because the segment sending rate during slow-start is always much higher than the available bandwidth when using the exponential growth in the slow-start mechanism. Moreover, they only allow one segment to be sent when starting or Chapter 2 Three TCP Implementations 23 restarting after packet losses. Compared with both T C P Reno and T C P R F C , T C P Vegas has much better congestion control strategies to control the traffic flow. The congestion avoidance actions of T C P Vegas are based on changes in the estimated amount of extra data in the network to maintain a steady data flow into the network without creating its own losses. Because of the fast retransmission and fast recovery mechanisms, T C P Reno has better retransmission strategies than T C P R F C . The time interval between sending a segment which is lost and retransmitting the lost segment in T C P Reno is much shorter than in T C P R F C . Moreover, the selective repeat A R Q in T C P Reno can significantly reduce the number of unneces-sary retransmissions. Therefore, the performance of T C P Reno is much better than that of T C P R F C . T C P Vegas further modifies the retransmission mechanism in T C P Reno by using a more accurate segment retransmission time. The time to retransmit a lost packet in T C P Vegas is much faster than that in both T C P Reno and T C P R F C which is the slowest one. In the round trip time estimations, T C P Vegas is more accurate than T C P Reno and T C P R F C because Vegas estimates R T T and updates the current R T O for every segment; however, the detailed calculations may cause excessive delays in low packet loss rate situations. Chapter 3 Heterogeneous Wireless and Wired Networking Environments This chapter provides an overview of wireless data communications systems and the Internet environment and described the relevant network simulation models. Two common commercially used wireless data communications networks, i.e., a broadband satellite network and a mobile data network, are studied here. Broadband satellite communications can provide communications coverage over a very wide area and interconnections for users at remote areas. Mobi l e cellular communications systems can provide wireless data communications with non-stationary hosts. 3.1 Broadband Satellite Communications Geostationary (GEO) satellites are currently used for satellite communications because of their fixed relay station. A satellite network can directly relay frames between two remote L A N s . The inherent advantages of frame relaying for L A N interconnections in a satellite network are simple access interface, statistical multiplexing capability, and relatively high access speed. However, because of the long distance between the satellite and the earth, the propagation delay between a source and a destination is about 0.27 sec for G E O satellite networks. If using the congestion avoidance and the error control methods in T C P , a sender must wait at least 0.54 sec before an acknowledgment for a transmitted segment is returned to the sender. Another factor that may degrade the performance of a satellite communication system is the possibility of a high bit error rate (BER) over a satellite link. In digital satellite systems [31] the nominal B E R should be in the order of 10"8 or lower for clear sky operation. If the B E R is worse than the B E R thresholds, usually about 10"3 or higher, the digital satellite l ink may lose 2 4 Chapter 3 Heterogeneous Wireless and Wired Networking Environments 25 modem synchronization and cease operations; therefore, the satellite system is unavailable for service at that point. A t the higher frequency bands employed by satellite systems, one of the main causes of B E R degradation is rain attenuation [38]. Attenuation caused by rain along a long transmission path is one of the dominant causes of signal fading in a space communication link. Rain drops absorb and scatter wave energy, resulting in a reduction in amplitude and randomness in the phase of the received signal, thus causing increased B E R . Therefore, rain attenuation seriously degrades the reliability and the performance of the satellite l inks. The probability of packet losses is directly related to the B E R and the packet length. If the current congestion control mechanism in T C P is used, the packet losses due to unreliable satellite l inks are also treated as indications of network congestion. Satellite Section / b \ \ Satellite Loss / Figure 3.1 Broadband Satellite Model 3.1.1 Broadband Satellite Model In order to take the effects of rain fades in satellite channels into consideration, packet losses over satellite links are modeled by a two-state Markov model with different B E R s for the clear sky and rainfall conditions. Figure 3.1 shows the broadband satellite model which is composed of a satellite delay node and a satellite loss node. In the satellite loss node, the two-Chapter 3 Heterogeneous Wireless and Wired Networking Environments 26 state Markov model is composed of a clear sky state and a rain fade state. State transition is assumed to happen only after the transmission of each packet over the satellite. Errors in adjacent bits within a packet are assumed to be independent of each other for both forward and backward channels. The packet errors in this model are assumed to be randomly distributed. The B E R in the clear sky state and in the rainfall fade state are represented by "BERC" and "BERR", respectively. The loss probabilities of a data packet and an acknowledgment in the clear sky condition are PCD and PCA , respectively. The loss probabilities of a data packet and an acknowledgment in the rain fade condition are PRD and PRA , respectively. The transition prob-ability from a clear sky state to a rain fade state is b , and the transition probability from a rain fade state to a clear sky state is g. The length of a data packet is L D . The length of an acknowl-edgment is LA PA is the average of the loss probability of an acknowledgment in the backward channel (see equation 2), and PD is the loss probability of data packet in the forward channel (see equation 1). Since the PA and PD are independent, the average packet loss probability in the satellite channel, P in equation (3) presents the probability of a packet not being acknowledged because either the packet or its acknowledgment is lost. (1) (2) Where PCD=\-(\-BERC) PRD=\-(\-BERR) P C A = l - ( l - BERC) P^I-H-BERR) Chapter 3 Heterogeneous Wireless and Wired Networking Environments 27 P = PD + (\-PD)PA V> 3.2 Mobile Cellular Communications A basic cellular system architecture consists of three main parts which are mobile units, a cell site, and a mobile telephone switching centre. The mobile units have a transceiver and an antenna system. The cell consists of a control unit (base station), radio cabinets, antennas, and data terminals, providing interface between the mobile units and the swithching centre. The switching centre is the central coordinating element for all cell sites (radio base station sites), providing interface among different telephone company zone offices and controlling call process-ing. Cellular mobile systems are characterized by the enhanced capabilities of wireless terminals to exchange signaling with the remainder of the network through the switching centers or base stations. Another main feature of the cellular mobile radio system is frequency reuse. In this frequency reuse system, users in different cells can simultaneously use the same frequency channel. The frequency reuse concept can be used in the time domain and the space domain. In the time domain, the frequency reuse occupies the same frequency in different time slots. In the space domain, the frequency reuse either can assign the same frequency in two different cells or repeatedly use the same frequency in the same general area in one system. The frequency reuse system can significantly improve the spectrum efficiency only i f the system is properly designed and implemented. T y p i c a l mobile wireless networks [6] [36] are composed of large numbers of base stations. Each base station has a coverage area and contains a channel which is used to establish, Chapter 3 Heterogeneous Wireless and Wired Networking Environments 28 maintain, update information, and terminate the connections of mobile hosts. The communica-tion between the base station and the mobile hosts can be based on a single C D M A / T D M A channel. Cellular Digi ta l Packet Data ( C D P D ) [33] is designed to operate as an extension of existing data communications networks as a peer multi-protocol, connectionless networks to existing data infrastructures. It provides a capability for 19.2 Kbps connectionless data packets layered on the top of the 30 K H z cellular phone channels and enables many data users to share one or more packetized data channels. Because the antenna height of a mobile unit is much lower than its typical surroundings and its carrier wavelength is much less than the sizes of the surrounding objects, multipath waves cause severe fading at various rates of fluctuation. The signal fluctuates in general in a range of about 40 dB, thus causing a relatively high packet loss rate. Besides packet losses due to severe signal fading, packets may also be lost during handoffs. Handoff is a process of automatically changing frequencies as a mobile unit moves into different cells; therefore, the conversation can be continued in new cells without reconnecting/redialing. Because of handoff processes, communications may pause and all the packets may be totally lost during the handoffs between cells. Therefore, i f the mobile unit moves fast, the rate of signal fluctuation is fast, and the probability of handoff is high, but the duration of fades is short. 3.2.1 Mobile Radio Model In a real wireless environment, wireless channels may randomly become unavailable for a certain period of time due to movement across cell boundaries (depending on how fast the mobile unit moves and how large the handoff region is). The loss characteristics of wireless channels is mainly bursty due to various fading effects. Therefore, a error model with uniform error probabil-ity used for modeling the loss characteristics of the wireless channel seems l ike ly to produce Chapter 3 Heterogeneous Wireless and Wired Networking Environments 29 inaccurate results. In order to capture the burstiness of wireless errors caused by the severe fading and the effects of mobility, packet losses over the wireless links are modeled by a three-state continuous-time Markov chain model based on the concept of the On-and-Off model [35]. The On-and-Off model assumes that the sources come on and go off for random intervals of time that are exponentially distributed with rate l / T o n and l /T o f f , respectively, where T o n is the average on time and T o f f is the average off time for a source. Figure 3.2 shows the mobile radio simulation model. Mobile Radio Section Figure 3.2 Mobile radio model This mobile radio model is composed of a Mobi le Delay node and a Mobi l e Loss node. The Mobi le Delay node presents the average end-to-end delay for every packet transmitting over the mobile radio links, including the propagation delay and the transmission delay. In the Mobi le Loss node, the three-state continuous-time Markov model is used to model the fast fading wireless channel which is composed of a handoff state and a normal state. In the handoff state, the end-to-end communica t ions is total ly stopped, wh ich means that the connect ion is Chapter 3 Heterogeneous Wireless and Wired Networking Environments 30 momentarily lost due to handoffs between cells. A l l packets in this state are lost. Because of the bursty nature of wireless channel errors caused by severe fading, a two-state continuous-time Markov chain model (a good state and a bad state) is incorporated within the normal state. In the normal state, any sent packets would be corrupted with a relatively high B E R when the channel stays in the bad state. When the channel stays in the good state within the normal state, any sent packets would be corrupted with a relatively low B E R . The packet errors in this model are also assumed to be randomly distributed. In this three-state model, each data communication alternates among states. The amount of time the channel spends in a state before making a transition into a different state is exponen-tially distributed with a defined mean. When the mobile unit moving within a cellular radio cell , the radio channel stays in the normal state. When the mobile unit moving within a handoff region, the radio channel stays in the handoff state. Figure 3.3 shows the occurrence of handoffs when a mobile unit moves into a different frequency radio cell. Figure 3.3 Occurrence of handoffs in a conventional cellular mobile system Assume that a mobile unit with a constant speed randomly initiates a call in a radio cell Chapter 3 Heterogeneous Wireless and Wired Networking Environments 31 with a defined radius and unidirectionally moves from one radio cell to another radio cell. When the mobile unit reaches a cell boundary, it needs to pass a handoff region before it enters into a new radio cell. A n average time of the mobile channel spending in a specified state (normal state or handoff state) is dependent on the size of the specified region and the mobile speed. In figure 3.3, lR' represents the radius of a radio cell, ' J ' represents the area of the handoff region, and T z ' represents the different frequency for a specified radio cell, where z = 1,2,3. The normal and handoff periods are independent and exponentially distributed with mean durations of 1/a sec and 1/b sec, respectively, where 1/a is represented as the average time between handoffs and 1/b is represented as the average time performing handoffs. The estimation of the 1/a and the 1/b are given by equations (7) and (8), respectively. In equation (4), y represents the probability of the mobile unit staying in a specified state at time t, C is the initial value, and k is the constant of proportionality. A t the initial point, t = 0, (for example, point A , right after the mobile unit entering into a new region), the probability of the mobile unit staying in the specified region (radio cells or handoff regions) is assumed to be 1. Equation (5) shows the amount of time the mobile unit takes driving from a initial point (x = 0) to x in meters within the specified region, where v is the mobile speed. When the mobile unit reaches to half way of the size of a specified region (e.g. d/2 in a the handoff region), the probabil-ity of the mobile unit staying in the specified region is assumed to be 0.5. Therefore, the mean duration of the exponential distribution can be estimated as equation (6). From both equations (7) and (8), 1/a and 1/b are inversely proportional to the speed of the mobile unit and directly propor-tional to radius of the mobile unit and the handoff region, respectively. (4) Chapter 3 Heterogeneous Wireless and Wired Networking Environments 32 t = x- (5) v Therefore, 1 0.5* fc~ln(0.5)xv 1 R W h e n x = 2fl, a ~ l n ( 0 . 5 ) x v (6) (7) When x = d, \ _ Q.5J (8) fc~ln(0.5)xv Within the normal state, the good and bad periods are also independent and exponentially distributed with means of 1/c sec and 1/d sec, respectively, where 1/d is the average duration of fades given by equation (9) and 1/c is the average duration between fades given by equation (11) in the steady state [36]. 1 _ J2n • Average duration of fades: ^ -  X T R (9) Level crossing rate: (no.of fades per sec) rate = -j= x TIR  i ^ x ^ D (10) 1 = 1 1 c rate d ( H ) In equation (9) and (10), (3, V , tR , and nR are wave number ( 2TC/A, ), speed of the mobile unit, the normalized duration of fades with respect to R (distance measured from the transmitter to the receiver), and the normalized level crossing rate with respect to R, respectively. Chapter 3 Heterogeneous Wireless and Wired Networking Environments 3.3 The Internet Environment 33 The Internet has grown rapidly by several orders of magnitude in size in recent years. T C P is the most widely used Internet transport protocol, operating above the Internet protocol (IP) at the network layer, which can be characterized as being either host limited or network limited. The host limited network model is assumed to have infinite buffer size; there is no packet loss in the network, but the queueing delay, in general, is relatively large. The network limited network model is assumed to have finite bandwidth and limited buffer size; thus, packet losses are mainly due to congestion in one or more bottleneck network nodes. Therefore, the probability of network congestion and the queueing delay for every packet over the Internet are the dominant factors on T C P throughput performance. In order to have a more practical Internet model, the characteristics of the Internet traffic and the average delay over the Internet are considered. Figure 3.4 illustrates the simulation Internet model, composing of a Internet congestion node and a Internet delay node. 3.3.1 Internet Congestion Model Since T C P packets make up roughly 80% of a l l wide-area network traffic over the Internet, the characteristics of the T C P traffic can be represented as the Internet traffic characteris-Chapter 3 Heterogeneous Wireless and Wired Networking Environments 34 tics. In general, network traffic varies significantly not only in traffic mix but also in connection characteristics both over time and between different sites [23]. The wide-area network applica-tions can be classified into two categories which are interactive traffic and bulk-transfer traffic. Interactive traffic is characterized as bidirectional small packet sizes ranging from 1 byte to 512 bytes. Over 90% of interactive conversations send fewer than 1000 packets and last less than a minute and a half. The ratio between the numbers of bytes sent by the originator and terminator of the interactive connection is about 1 to 20. Bulk-transfer applications are not strongly bidirec-tional and are request-response in nature. In order to generate a random but realistic sequence of traditional internetwork conversa-tions and data traffic for a set of sites, the preferred internetwork loading is bursty traffic which can be achieved by using different T C P traffic load models, as illustrated in figure 3.5. Without enforced synchronization, the simulation test scripts can generate a natural variation of traffic load during the long simulation duration. The author of [22] [24] found that most arrivals of T C P application conversations appear well-modeled as Poisson processes; however, it is not the case for all T C P data inter-packet arrivals. Most of the bytes transmitted by the T C P applications are sent by FTP, SMTP, N N T P , T E L N E T , and R L O G I N . In the Internet traffic model used in the simulation (see figure 3.5), the only T C P applica-tions included are FTP, SMTP, N N T P , T E L N E T , and R L O G I N . The arrivals of new T C P connec-tion requests are modeled as time-varying Poisson processes with a specified site and time-of-day dependent rate. The TCP/ IP wide-area characteristic library (tcplib) [23] is used to determine the characteristics of these T C P applications such as the duration of a specified T C P connection and the data size per connection. For interactive applications including T E L N E T and R L O G I N , the length of conversation in mill isecond unit in the source is determined by the telnet_duration function in the tcplib. During the conversation, single byte packets are sent from the sender until Chapter 3 Heterogeneous Wireless and Wired Networking Environments 35 the time duration expires. The interarrivals between two source packets in the sender are obtained from the telnet_interarrival function in the tcplib. After receiving a single byte packet from the sender, the destination responds with a packet which size is determined by the telnet_pktsize function in the tcplib. 10 Mbps 45 Mbps Figure 3.5 TCP application sources and Internet traffic model For the bulk-transfer applications including FTP, N N T P and S M T P , the number of items sent per conversation in the sender and the number of bytes sent per item are determine by the specified functions in the tcplib. A source running F T P applications is simulated by using the ftp_nitems function in the tcpl ib to decide how many items to send. For each i tem, the ftp_itemsize function is used to determine the size of each item. Since the packet arrival process for a bulk-transfer data connection is largely determined by network factors such as the available bandwidth, congestion, and details of the transport-protocol congestion control algorithms, the distribution of the burst data sizes in bytes is quite heavy-tailed. Therefore, the packet interarrival times of these bulk-transfer are characterized by the Pareto Distr ibut ion. The handshaking between item transfers uses the ftp_ctlsize function in the tcplib to get the size of the handshake Chapter 3 Heterogeneous Wireless and Wired Networking Environments 36 packets. N N T P and S M T P applications are simulated in a similar manner as F T P applications; however, each S M T P and N N T P conversation transfers only one item. Since the probability of the packet loss due to network congestion depends on buffer, size, the bottleneck router in the Internet model is characterized as being network l imited with a assumed buffer size of 30 Kbytes (a typical router buffer size). It is modeled as a store and forward queueing process model accepting packets from any number of sources and autono-mously forwards them to a single destination module. The packet's service time may vary from packet to packet and is computed by dividing packet length, measured in bits, by the data rate of the T3 link. The conversation arrival rates of FTP, N N T P , SMTP, Telnet, and Rlogin are modeled as Poisson processes with rates of 42.3, 5.76, 3.27, 32.7, and 30.7 conservations per sec, respec-tively (measured arrival rate of U C B conversations at 10 am. [23]). Over a series of simulations, the probabil i ty of the network congestion in the bottleneck router in the Internet model is measured to be 0.05 ±0.01 . 3.3.2 Internet delay model In the Internet, packets are transferred as datagrams from their sources node to their destination node by the defined IP routing tables. Each datagram is handled independently from all other datagrams. This means that IP datagrams can be delivered out of order. If a source sends two consecutive datagrams to the same destination, for example packets A and B , each is routed independently and can take different routes, with packet B possibly arriving at the destination before packet A . The observations of the end-to-end delay in the Internet [32] are that the round trip time varies substantially even over short periods of times and most of the losses occur one packet at a time. After a period of the dynamic changes of the round trip time, a step change behavior in Chapter 3 Heterogeneous Wireless and Wired Networking Environments 37 round trip time may remain effective for several seconds at a time, with only minor variations during this steady time. Therefore, a constant, exponential, or normal probability density function may either overestimate the end-to-end delay in the Internet or underestimate it to produce an inaccurate delay model. A better Internet delay model can be modeled by a sum of two functions [6], one of which is constant and the other is a random variable with an Erlang distribution. The constant term, which is the minimum network delay, can represent the average network delay in the case that no network congestion occurs and the queueing delay in each intermediate node is small . The variation of the network delay for a packet is dependent on the arguments of the Erlang distribution, which are a mean-outcome and an order of the distribution, as shown in equation (12). Equation (13) and (14) show the mean-outcome and the variance of the Erlang distribution. f ( x ) _ h a \ n - l e - a X n ) / ( n - l ) \ * 0 > o l JMv) " | Q | ( 1 2 ) E(x) = naX (13) 8 / = na1 (14) In equation (12), (13), and (14), n and a are the order and constant, respectively, in the condition of n > 0 when a > 0. In order to capture most Internet end-to-end delay behaviors, a two-state Markvo chain model is used to model the Internet delay model (see figure 3.6). In state 1, a larger mean value of the Erlang distribution is used to generate a relatively long end-to-end delay. In state 0, a smaller mean value of the Erlang distribution is used to generate a relatively short end-to-end delay. The transition probability from state 0 to state 1 is SOI, and the transition probability from state 1 to Chapter 3 Heterogeneous Wireless and Wired Networking Environments 38 state 0 is 510. Transition happens only after the transmission of each packet over the Internet. Based on the measurement results in [6] and [32], 96% of the packets stay in State 0. When a packet is in State 0, the packet has a shorter end-to-end Internet delay with minor variations. On the other hand, 0.04% of the packets stay in State 1. When a packet is in State 1, the packet has a longer end-to-end Internet delay, producing a few sharp rises within a few seconds. B y matching the means and the standard deviations of the first three sets of the experimen-tal results in [32], Erlang distributions with an order of 2 is used for both states, the same order as suggested in [6]. A mean delay of 0.001 sec is used in State 0. 75% of the packets stays in State 1 with a mean delay of 0.01 sec and 25% of the packets stays in State 1 with a mean delay of 0.1 sec. The constant delay of 0.01 sec is used for the constant function in the delay model. sOl slO Figure 3.6 Internet delay model Chapter 4 Analysis of TCP in a Satellite Environment In this chapter, the performance of three TCP implementations (TCP Vegas, TCP Reno, and TCP RFC) in an environment of a broadband satellite interworking with the Internet (BSII) is studied. Optimization of some key TCP parameters for the TCP implementations in the BSII environment is presented. 4.1 Satellite Network Interconnection A schematic topology of a broadband satellite wide area network interworked with the Internet is given in Figure 4.1. A satellite channel is assumed to function as a transparent path between workstations on a remote LAN and its Internet gateway, communicating with other workstations on a different LAN over the Internet. Satellite Figure 4.1 Satellite interconnected network topology Propagation delay and transmission delay are two delay elements in the satellite channel. A gateway connects the satellite channel to the Internet. The TCP layer in one of the workstations 39 Chapter 4 Analysis of TCP in a Satellite Environment 40 accepts bulk data files and breaks them up into smaller segments with a fixed maximum segment length for transmissions. The IP layer accepts the segments from TCP, appends an IP header to each segment, and sends each piece as a datagram. The datagrams are sent through the lower layer of the L A N s and passed through the Internet. A t the destination, T C P reassembles all the segments according to the sequence number assigned to each segment and recovers the original file. Aspect of communications related to the data link and the physical layer are not addressed in this thesis. In order to maximize T C P performance over the satellite network interconnected with the Internet, a number of T C P dependent parameters such as the maximum transmit window size ( M T W S ) , the maximum A C K delay ( M A D ) , and the Maximum segment size (MSS) , etc., need to be jointly optimized under different conditions. 4.2 Simulation Model To perform a detailed analysis of the behavior of different T C P implementations in the BSI I environment, a sophisticated queueing network model for a typical Internet is developed. Packets are transferred from a source node to a destination node along paths and routers specified by IP routing tables. Figure 4.2 shows the simulation model for the BSII environment, including a number of nodes. The simulation model is divided into two sections which are the broadband satellite section and the Internet section. The gateway connects the broadband satellite section to the Internet section. The satellite section models the characteristics of the satellite channel described in section 3.1. The Internet section models the characteristics of the Internet described in section 3.3. In the simulation model, data packets are sent from the source node to the destina-tion node, and the corresponding acknowledgment packets are returned immediately from the destination node to the source node after a fixed A C K delay. Passing through the Internet section and the broadband satellite section, the data packets and A C K are delayed and may be lost due to congestion or transmission errors in the media. Chapter 4 Analysis of TCP in a Satellite Environment 41 Broadband Satellite Section Internet Section Figure 4.2 The simulation network model 4.3 Simulation Assumptions and Parameters The different T C P implementations described in section 2.2, gateways, routers, the satellite section, the Internet section, the source and the destination are modeled by O P N E T , a simulation tool. Since the big window option in [18] [21] can give a method of expanding the standard 64 Kbyte maximum window size to a 1 gigabyte allowable receive window size, the big window option is assumed to be used in all T C P end-to-end connections. File transfer is the only application considered for sending data from the source node through the Internet to the destina-tion via a satellite channel. Fi le transfer is performed over a logical connection that is opened between the transport layer at the two end nodes. Duplex transmission l inks are used in the simulation model. The satellite links are assumed to be lossy digital pipes subjected to transmis-sion and propagation delays. The gateway and the routers function as a store-and-forward packet switching nodes with infinite buffer sizes in which packets are only subjected to queueing delay. The only source of errors in the Internet section is the packet losses due to network congestion. Because of the serious effects of rain attenuation on a satellite link, the only packet Chapter 4 Analysis of TCP in a Satellite Environment 42 losses in the satellite channel are assumed due to rain fade conditions. These corrupted segments are detected by the checksum in the T C P layer and are recovered by the retransmissions. In the simulations several variable parameters including the packet loss rate ( P L R ) in the satellite channel, the average end-to-end (ETE) delay measured from the time the packet leaves the T C P layer in the source node to the time the packet is forwarded to the application in the destination, and the key T C P parameters including the M T W S , the M A D , and the M S S are considered. Table 4.1 summarizes the fixed simulation parameters for all the simulations in this chapter. Table 4.1 Fixed simulation parameters for the simulation model Simulation Parameters Header for IP datagram 20 bytes Header for TCP segment with a big window option 22 bytes Data rate of the LANs 10 Mbits/sec Data rate of the Internet 45 Mbits/sec Data transferred per run 30 Mbytes *The average gain of the smooth RTT estimation 0.125 *The deviation gain of the smooth RTT estimation 0.25 The data rate of the satellite channel 1.544 Mbits/sec The average one-way satellite delay 0.28 sec The average congestion probability for the Internet model 0.04 The average end-to-end Internet delay 0.018 sec The constant term of the Internet delay distribution 0.01 sec The mean of the Erlang distribution for state 0 in the Internet 0.001 sec The mean of the Erlang distribution for state 1 in the Internet 0.01 sec and 0.1 sec * The average gain and the deviation gain of the smooth RTT estimation is used for estimating the retransmission time-out for the specified TCP segments. 4.4 Impact of Packet Loss Rate on TCP Throughput This section presents the effects of the P L R due to the unreliable satellite l inks on the throughput of different T C P implementations. Satellite links are usually designed with low clear Chapter 4 Analysis of TCP in a Satellite Environment 43 sky B E R ; however, B E R and hence P L R cou ld increase substantially dur ing rain fades. Frequency and duration of rain fades depend on the climatic region in which the earth station is located. Therefore, to determine the effects of the P L R in the satellite link on T C P performance under different conditions, the P L R in the rain fade state and the state transition probabilities (STPs)are varied in a series of simulations. A l l the key T C P parameters are fixed with values as suggested in [21] [17]. The thresholds ( a and p ) in T C P Vegas are assumed to be 1 and 3. The R T T coefficient deviation is assumed to be 4. The average E T E delay is assumed to be 0.3 sec. Figure 4.3 compares the throughput performance of different T C P implementations as a function of the P L R in the satellite channe. In figure 4.3, state transition probability (STP) from clear sky state to rain fade state is assumed to be 0.01 (a upper bound percentage of the measured rain fade level [39]) and STP from clear sky state to rain fade state is assumed to be 0.05. 120 1 1 1 1 MTWS = 136 Kbytes 100 MSS = 512 Bytes MAD = 10 msec £ so RTT Coef. = 4 £ -3 Q. J 3 4 0 6 0 \ \ V e g a s 3 O H 20 \ R o n o y R F C "~-" — • • i i i O 0.01 0.02 0.03 0.04 0.05 0.06 P a c k e t L o s s R a t e i n t h e S a t e l l i t e L i n k Figure 4.3 Throughput of different TCP versions vs. the PLR in the satellite link From figure 4.3, because the slow-start mechanism treats packet losses over the satellite channel as indications of congestion, the throughput of the three different T C P implementations decreases significantly when the P L R in the satellite l ink increases. Because of the dynamic fluctuation of the congestion window, the long delay satellite channel cannot be fully utilized even Chapter 4 Analysis of TCP in a Satellite Environment 44 with the implementation of the big window option. In this satellite channel, T C P Vegas gives approximately four times better throughput than T C P Reno and six times better throughput than T C P R F C for Packet loss rates (PLRs) in the satellite links less than 0.03. Beyond the P L R of 0.03, the throughput of T C P Vegas comes closer to the throughput of both T C P Reno and T C P R F C . The throughput of both T C P Reno and T C P R F C do not change much when the P L R is increased beyond 0.02. Overall, the P L R has much greater effect on the throughput of T C P Vegas over the satellite channel than the other T C P implementations, and T C P Vegas has much better throughput than the others under the whole range of the P L R s . Figure 4.4 plots the congestion window response for different T C P implementations versus time under a clear sky condition. It also shows a more detailed picture of the behavior of the congestion window for the T C P implementations, responding to the packet losses mainly due to the network congestion. Figure 4.5 and 4.6, respectively, plot the congestion window response for T C P Reno and T C P Vegas versus time under different STPs with a B E R of 5x10~ 6 in the rain fade state. They show the congestion window's behavior for both T C P versions, responding to the packet losses not only due to the congestion but also due to unreliable satellite channel. For every packet loss, the congestion window size is reduced linearly by almost one M S S in T C P Vegas, by almost one half in T C P Reno, and to one M S S in T C P R F C . In figure 4.4, because of its congestion avoidance approaches, T C P R F C has more sharp drops of the congestion window to one segment size than the other T C P implementations; whereas T C P Vegas has much less frequent drops of the congestion window than the others. However, because of its increasingly large congestion window and the exponential back-off algorithm, T C P Vegas seems to have longer communication pauses than the other two implemen-tations i f packet losses are detected by the time-out mechanism. This phenomenon appears as the flat and empty region of the curve indicated by "backoff period". Because the congestion Chapter 4 Analysis of TCP in a Satellite Environment 45 windows in both TCP RFC and TCP Reno are limited by Internet congestion, their throughputs are lower than the throughput of TCP Vegas even when there is no packet loss over the satellite channel. (xle+04) 2.51 Figure 4.4 Congestion window response in different TCP implementations In both figure 4.5 and 4.6, '*' represents the loss of a data packet and 'o' represents the loss of an acknowledgment frame. The STP from clear sky state to rain fade state is referred as 'b', and the STP from rain fade state to clear state is referred as 'g'. When b increases, the average duration of clear sky (between rain fades) is accordingly decreased, thus the satellite link Chapter 4 Analysis of TCP in a Satellite Environment has a relatively high probability of staying in the rain fade state. 46 Figure 4.5 The congestion window response in TCP Reno with different STPs Figure 4.5 shows how the congestion window responses in T C P Reno when data packets are lost within a full transmit window size under different STPs. In figure 4.5 (a), most of the retransmissions rely on the F R F R algorithm. A t a higher b and a lower g (see figure 4.5 (b)), the average duration of each rain fade event (1/g) is relatively long; therefore, the probability of several corrupted data packets within a full transit window size is relatively high. The recovery of the connection in T C P Reno from a channel interruption is not fast enough to expand the current congestion window size before the next channel interruption arrival. Therefore, most of the retransmissions rely on the R T O mechanism and thus the congestion window size is always limited by this mechanism. Chapter 4 Analysis of TCP in a Satellite Environment 47 Figure 4.6 The congestion window response in TCP Vegas with different STPs Figure 4.6 shows the congestion window response in T C P Vegas when there are several corrupted data packets within a full window size. Since T C P Vegas has a better error recovery mechanism in response to the numerous lost packets over the satellite channel, the congestion window size does not dynamical ly fluctuate even when b is increased and g is decreased. However, the current congestion window size is still limited by the increasing number of retrans-missions. Because these channel interruptions due to the unreliable satellite link totally mislead the flow control mechanism in T C P Vegas, the congestion window is increased until when the fine-grain time-out mechanism in T C P Vegas fail to recognize a lost segment. Coarse-grain time-out as in Reno is employed. In figure 4.6 (b), because of the larger congestion window and many lost packets within a Chapter 4 Analysis of TCP in a Satellite Environment 48 fu l l transmit window size, the exponential back-off pol icy is extended and causes a long communication pause before the channel is recovered. Compared with T C P Reno and T C P R F C , T C P Vegas still has much better congestion avoidance to prevent losses due to congestion and has a much faster error recovery mechanism to quickly recover from any losses over the satellite channel. Therefore, the throughput performance of T C P Vegas, overall, is much better than the others for different P L R s over the satellite links. 4.5 Impact of End-to-End Delay on TCP Throughput This section presents the effects of the E T E delay on T C P throughput. Because the satellite link does not have the high bandwidth of wired links and have higher latencies, there is substantially bandwidth changes between wireless and wired portions of the communications path. The queueing delay in the satellite system, which is the sum of the queueing delay in satellite modems and gateways between the wireless and wired portion of the communication path, is a critical element for this interconnection. A n exponential distribution with a variable mean is used to represent the variations in queueing delay of the satellite system. In this section, all the network elements and T C P parameters are fixed and the same as those in section 4.6, but the average queueing delay in the satellite system is varied, and the average E T E delay is measured accordingly. Figure 4.7 compares the throughput performance of different T C P implementations as a function of the average E T E delay under a 0.01 P L R in the satellite channel and a clear sky condition. Figure 4.8 plots the sequence number versus time for the T C P data transfer over the satellite channel with a P L R of 0.001 versus time. Figure 4.7 shows that the throughput efficiency of all T C P versions decreases linearly as the E T E delay increases under both clear sky and a P L R of 0.001 conditions. With increasing E T E delay, the sender must wait for a longer round trip delay for an A C K to be returned before it can send or retransmit packets into the network. As a result, the data sending rate in the sender is Chapter 4 Analysis of TCP in a Satellite Environment 49 significantly reduced and the error recovery period is lengthened, thus causing T C P throughput degradation. The E T E delay seems to have less effect on the throughput of T C P R F C than on that of both T C P Vegas and T C P Reno. Because of the faster error recovery and the better congestion avoidance mechanisms, T C P Vegas still has higher throughput than the other T C P versions over the whole range of the E T E delay even when the bandwidth delay product is large. x> 3 a. JS OD 3 O Vegas 0.001 PLR Clear Sky Condition Vegas 0.45 O.S O.S5 Average end—to—end delay (sec) Figure 4.7 Throughput of different TCP versions vs. average ETE delay Figure 4.8 shows a more detailed picture of the behavior of the E T E connection for differ-ent T C P implementations. The figure also shows the comparison of sequence number progression in a connection using T C P Vegas, T C P Reno, and T C P R F C implementations for different E T E delays. B y increasing the E T E delay from 0.35 sec to 0.7 sec, there is approximately 40% reduction in the rate of the sequence number progression in all T C P implementations. A s the E T E delay increases, the communication pauses become more often and each of the communica-tion pause lasts much longer. This phenomenon appears as the empty regions of the curves. T C P Vegas maintains a relatively high and consistent throughput even at the E T E delay of 0.7 sec; however, both T C P Reno and T C P R F C unneccessarily invokes congestion control procedures Chapter 4 Analysis of TCP in a Satellite Environment many times during the data connection. 50 Sequence Number (xle+06) 7iOo no" 120" Sequence Number (xle+06) 130 ~m 150 Time (sec) T C P Reno 100 110 120 130 ..jsequerjGe JSIurnfcer i x j e+95)„_ , 150 time (sec) T C P R F C 140 150 time <seo) Average End-to-end delay = 0.35 sec Sequence Number (xle+06) T C P Vegas ISO tiine (seo) Sequence Number (xle+06) T C P Reno 00 110 120 Sequence Number (xle+05) 150 time (sec) T C P R F C Average End-to-end delay = 0.7 sec ISO time <sec) Figure 4.8 Sequence number for data transfer over the satellite channel with a 0.001 PLR 4.6 Impact of TCP Parameters on TCP Throughput This section analyzes the effects of TCP parameters including the MTWS, the MAD, and the MSS on the throughput of the different TCP implementations over the satellite channel described in section 3.1. The MTWS and the MAD mainly control the rate of data transmission. The MSS not only controls the number of data bytes in transit per round trip time but also has a great effect on the PLR over the satellite links. These parameters strongly affect TCP throughput. If the values of these TCP parameters are too large or too small, the bandwidth efficiency of the Chapter 4 Analysis of TCP in a Satellite Environment 51 system may fluctuate over a wide range, thus causing T C P throughput inefficiency. Therefore, a detailed evaluation of these parameters is presented here. In this section, the thresholds ( a and (3 ) in T C P Vegas are assumed to be 1 and 3 [26], the R T T deviation coefficient is assumed to be 4 [17], b is assumed to be 0.01, and g is assumed to be 0.05. 4.6.1 Effects of Maximum Transmit Window Size Many pervious studies in satellite interconnections [3] [5] [6] indicated that the M T W S strongly affects T C P throughput. In this subsection, the effects of the M T W S on the throughput of different T C P implementations are presented. A l l network components and T C P parameters are fixed for performance evaluation by only varying the M T W S . The M S S is assumed to be 512 bytes. The M A D is assumed to be 10 msec. The throughput of T C P R F C , T C P Reno, and T C P Vegas as a function of the M T W S for different P L R s under a 0.3 second average E T E delay condition are illustrated in figure 4.9, figure 4.10, and figure 4.11, respectively. Figure 4.12 and 4.13 show the throughput performance of T C P Reno and T C P Vegas as a function of the M T W S for different E T E delays under a clear sky condition. B y using the big window option [21], all T C P implementations can fully utilize the long satellite channel under no packet loss condition; however, the capability of this option is limited by the limited Internet resources. At a relatively low P L R , T C P throughput cannot be increased further and eventually drops when the M T W S increases beyond a certain value, indicating that the network resources are overloaded. Due to the strategies of the existing retransmission and congestion control mechanisms, al l T C P implementations unneccessarily reduce the current congestion window size and over-activate the exponential back-off algorithm particularly in a relatively high P L R environment. A s a result, their throughput performances are degraded and the benefits of the larger transmit window are nullified. Chapter 4 Analysis of TCP in a Satellite Environment 52 In figure 4.9, a M T W S greater than 16 Kbytes has nearly no influence on the throughput performance of T C P R F C for different P L R s because the Go-Back-N A R Q strategies is used for retransmissions. Beyond a certain M T W S for a specified P L R in the satellite link, T C P through-put seems to be constant over a wide range of the M T W S s . Therefore, there is no need for big window option in T C P R F C . Due to the activity of the slow-start mechanism at high P L R s , the congestion window size always be reset to one M S S , thus degrading extremely T C P throughput. c l e a r s k y P a c k e t l o s s r a t e = 0 . 0 0 1 P a c k e t l o s s r a t e = 0 . 0 1 P a c k e t l o s s r a t e = 0 . 0 5 6 0 8 0 1 0 O T r a n s m i t W i n d o w S i z e ( K b y t e s ) Figure 4.9 Throughput of TCP RFC vs. M W T S for different PLRs Figure 4.10-4.13 show that the M T W S has greater influence on the throughput of T C P Vegas than on that of T C P Reno for different P L R s and different E T E delays. The big window option works well and greatly improves the throughput for both T C P implementations only at a relatively low P L R s and long E T E delays. Beyond an upper limit of the M T W S , the throughput of both T C P Reno and T C P Vegas decreases sharply for different P L R s and different E T E delays. It is because a bigger M T W S causes the sender to overflow the buffer in the Internet and creates more packet losses. A t a higher P L R or a longer E T E delay under clear sky condition, increasing the M T W S beyond certain values seems to have no influence, particular in T C P Reno, on the throughput performance. Chapter 4 Analysis of TCP in a Satellite Environment 53 Chapter 4 Analysis of TCP in a Satellite Environment 54 1 4 0 T r a n s m i t W i n d o w S i z e ( K b y t e s ) Figure 4.13 Throughput of TCP Vegas vs. M T W S for different E T E delays The optimal maximum transmit window sizes (MTWSs) and their corresponding maximum throughput for different TCP implementations under different PLR conditions are summarized in Table 4.2. It seems that a bigger MTWS is suitable for TCP Reno and TCP RFC only at a relatively low PLR. At a high PLR, a much smaller MTWS seems to give a slightly better throughput for both TCP Reno and TCP RFC. TCP Vegas performs well not only at a relatively low PLR but also at a relatively high PLR. At a PLR of 0.001, TCP Vegas has approxi-mately three times better throughput than TCP Reno and approximately thirteen times better throughput than TCP RFC. At the PLR of 0.01, TCP Vegas still has nearly six times better throughput than TCP Reno and nearly eight times better throughput than TCP RFC. Packet losses over the Internet cause the sender's congestion window size to dynamically fluctuate in both TCP Reno and TCP RFC, resulting in a small average transmit window size. Therefore, the throughput of both TCP implementation is limited by the Internet congestion. The use of a bigger MTWS in both TCP versions results in increasing packet losses in the Internet due to overload at the limited network buffers, and hence reduces their throughput performances. With a better congestion avoidance mechanism, TCP Vegas can take an advantage of the big window option to send more data packets into the network without easily overflowing the Chapter 4 Analysis of TCP in a Satellite Environment 55 network buffer. With a faster error recovery mechanism, T C P Vegas can overcome the trade-off of the bigger M T W S to work reasonably wel l even at a high P L R . T C P Vegas with a bigger M T W S , overall, has much better throughput than the other T C P implementations in the H W W network; whereas, T C P R F C has worse throughput than the others under most conditions. Table 4.2 The optimal MTWSs for different TCP versions under different conditions TCP Vegas TCP Reno T C P R F C M T W S for the clear sky condi-tion 128 KB 128 KB 24 KB Throughput (Kbytes/sec) 78.52 130.15 11.89 M T W S at a PLR of 0.001 88 KB 64KB 16KB Throughput (Kbytes/sec) 108.79 34.19 8.27 M T W S at a PLR of 0.01 64KB 16KB 8 KB Throughput (Kbytes/sec) 42.76 7.26 4.39 4.6.2 Effects of Maximum ACK Delay The sending rate and the error recovery period in the transmitter depend on the frequency of the A C K frames received. If the M A D is too small to allow out of sequence packet arrivals, unnecessary retransmissions w i l l be increased; as a result, the data packet sending rate w i l l be decreased, thus wasting a large amount of bandwidth. If the M A D is too large, the sender congestion window wi l l grow too slowly, also causing low throughput performance. This subsec-tion mainly studies the effects of the M A D on the throughput of different T C P implementations and presents the optimal maximum A C K delays ( M A D s ) for different conditions. A l l the other parameters are fixed, but the M A D is varied for different P L R s and different average E T E delays. The M T W S is assumed to be 128 Kbytes. The segment size is assumed to be 512 bytes. The throughput of T C P R F C , T C P Reno, and T C P Vegas as a function of the M A D for different P L R s Chapter 4 Analysis of TCP in a Satellite Environment 56 under a 0.3 second average ETE delay condition are illustrated in figure 4.14, figure 4.15, and figure 4.16, respectively. From figure 4.14, when the MAD is increased, the throughput of TCP RFC is gradually decreased and becomes constant. A smaller MAD gives higher throughput for TCP RFC at low PLRs. However, it seems that the MAD has a smaller effect on the throughput of TCP RFC at high PLRs. Figure 4.15 and 4.16 show that the throughput of TCP Reno and TCP Vegas increases as the MAD increases to a certain value for different PLRs, beyond which the throughput of both TCP versions is decreased. TCP Vegas with a bigger MAD seems to improve it throughput only for clear sky condition. Under high packet loss conditions, the MAD does not have great effects on the throughput of TCP Reno and TCP Vegas. Figure 4.17 and 4.18 show the throughput of TCP Reno and TCP Vegas as a function of the MAD for different average ETE delays under a clear sky condition. At a smaller average ETE delay, the throughput of both TCP versions increases to a certain MAD and then gradually decreases. At a longer average ETE delay, the much smaller MAD seems to improve the through-put for both TCP Reno and TCP Vegas because a smaller MAD can reduce the error recovery period especially for the much longer bandwidth-delay products. M a x i m u m A C K D e l a y ( s e c ) Figure 4.14 Throughput of TCP RFC vs. M A D for different PLRs Chapter 4 Analysis of TCP in a Satellite Environment « s o r a t e •= O . O O I - p a c k e t l o s s r a t e = 0 . 0 1 . z p a c k e t l o e i s r a t e = 0 . 0 5 1 1 • 1 1 M a x i m u m A C K D e l a y ( s e c ) Figure 4.15 Throughput of TCP Reno vs. M A D for different PLRs &3 t o o c l e a r s k y p a c k e t l o s s r a t e = 0 . 0 1 p a c k e t l o s s r a t e = O . O S O . O I S 0 . 0 2 0 . 0 2 5 M a x i m u m A C K D e l a y ( s e c ) Figure 4.16 Throughput of TCP Vegas vs. M A D for different PLRs e t o = 0 . - 4 5 s e c e t e = 0 . 6 5 s e c e t e = 0 . 8 5 s e c M a x i m u m A C K D e l a y ( s e c ) Figure 4.17 Throughput of TCP Reno vs. M A D for different E T E delays Chapter 4 Analysis of TCP in a Satellite Environment 58 M a x i m u m A C K D e l a y ( s e c ) Figure 4.18 Throughput of TCP Vegas vs. M A D for different E T E delays Table 4.3 The optimal maximum A C K delays for different TCP versions under different conditions T C P Vegas T C P Reno T C P R F C M A D under a clear sky condi-tion 20 msec 2 msec 0.5 msec Throughput (Kbytes/sec) 145.47 66.76 17.67 M A D at a PLR of 0.001 10 msec 1 msec 0.1 msec Throughput (Kbytes/sec) 102.55 53.48 15.12 M A D at a PLR of 0.01 10 msec 1 msec 0.1 msec Throughput (Kbytes/sec) 46.54 15.57 5.82 The optimal M A D s and the corresponding throughput for different T C P implementations under different conditions are summarized in Table 4 .3. A t different P L R s , a relatively long M A D seems to give higher throughput for T C P Vegas; however, a relatively short M A D seems to give higher throughput for both T C P Reno and T C P R F C . Because T C P Vegas uses its finer-grained timer to calculate a more accurate RTT estimation and update the current R T O for every segment, the accurate information in the returning A C K is very important. If the M A D is not large enough to wait for all out of sequence packet arrivals in the receiver, T C P Vegas in the Chapter 4 Analysis of TCP in a Satellite Environment 59 sender w i l l receive the incorrect information from the returning A C K . A s a result, it unneccessar-ily retransmits all the late arrival data packets in the receiver. Because of these misinterpretations, T C P Vegas may easily overflow buffers, thus causing more losses for the other connections which are sharing the same resources. Because the retransmission of the lost packets in T C P Reno relies totally on the third duplicated A C K received and the retransmission of that in T C P R F C relies only on the R T O , the information in the returning A C K in both T C P s does not need to be as accurate as that in T C P Vegas. Moreover, the rate of the exponential growth of the congestion window size after a loss in both T C P implementations depends on the number of A C K frames received. Therefore, T C P Reno and T C P R F C prefer to use a relatively short M A D in order to quickly increase the data sending rate after a packet loss. A t high P L R s , all three T C P implementations prefer to have a shorter M A D in order to have much faster error recovery. However, at a low P L R , the T C P implementations prefer to have a relatively larger M A D in order to prevent the buffers from overflowing. A t different P L R s in the satellite links and average E T E delays, the M A D has greater influence on the throughput of T C P Vegas than on that of T C P Reno and T C P R F C ; whereas the M A D has the least influence on the throughput of T C P R F C . Overall, T C P Vegas has better throughput performance than the others over the whole range of the M A D under most conditions. 4.6.3 Effects of Maximum Segment Size In order to obtain the maximum throughput performance for the T C P connection, the sender needs to send enough segments which should be as large as possible to f i l l up this long satellite pipe. Typically, the M S S has a slight effect on T C P throughput in the wireline networks because of the almost negligible packet lost rate due to the media. However, under a rain fade Chapter 4 Analysis of TCP in a Satellite Environment 60 condition, the average bit error rate is relatively high in a satellite channel, resulting in many corrupted packets. Since the P L R are directly depended on a specified B E R and a specified segment size, the net effect of the segment size for the specific B E R is to change the packet error rate, which effects on T C P have been considered in section 4.4. Equation (10) shows the relation-ship among the packet error rate, the segment size, and the B E R . If the segment size increases at a specified B E R , the P L R is almost l inearly increased. Therefore, a smaller segment size is suggested for a relatively high B E R environment in order to decrease the P L R [6] [15]. PacketErrorRate = 1 - (1 - BER)SesmentSize ( 1 0 ) However, it is not always true in the case of communications over a satellite channel with a relatively high STPs between clear sky and rain fade states. Typically, the satellite channel stays in the clear sky state for a while before the next channel interruption occurs. A t a specified B E R , the P L R in the satellite channel is not only dependent on the M S S but also on b and g (see equations 1 and 2 in section 3.1.1). A t a higher b and a lower g, the probability of the satellite link staying in the rain fade state is relatively high. A larger segment size has a higher probability of error per packet than a smaller segment size particular in the rain fade state. Therefore, a larger segment size causes a relatively high P L R in the satellite link. However, at a reasonably low b and a reasonably high g, the probability of the satellite l ink staying in the rain fade state is relatively low. A larger segment size can carry more data information for every round trip time before the next channel interruption occurs. Therefore, a larger segment size may improve T C P throughput. This subsection investigates the effects of the M S S on the throughput of different T C P implementations in the H W W environment over a range of the P L R in the forward satellite channel (PD in equation 1). A l l simulation parameters are fixed, but the P L R in the satellite links and the M S S are varied. The M T W S s in T C P Vegas and T C P Reno are 128 Kbytes and the Chapter 4 Analysis of TCP in a Satellite Environment 61 M T W S in T C P R F C is 16 Kbytes. The M A D s in T C P Vegas, T C P Reno, and T C P R F C are assumed to be 10 msec, 1 msec, and 0.1 msec, respectively. Figure 4.18, figure 4.19 and figure 4.20 show the throughput performance of T C P R F C , T C P Reno, and T C P Vegas, respectively, as a function of the P L R in the satellite channel for different maximum segment sizes (MSSs). F rom those figures, the throughput of the different T C P implementations is increased when the M S S is increased from 512 bytes to 1458 bytes over a whole range of the P L R in the satellite links. It is because a larger segment size can carry more data information for every round trip time even through there is a trade-off of a relatively high probability of error per packet. However, the throughput of all T C P implementations is decreased when the M S S is increased from 1458 bytes to 1460 bytes because of the increase of the probability of IP fragmentation, as explained later, and a relatively high probability of error per packet. The M S S seems to have no effect on the throughput of T C P R F C beyond a P L R of 0.05. The effects of the M S S on the throughput of T C P Vegas is greater than that on the throughput of T C P Reno over the whole range of P L R s . A t a relatively high P L R in the satellite channel, for example, P L R > 0.03, the through-put of T C P Reno and T C P Vegas with a M S S of 1460 bytes is slightly better than that with a M S S of 512 bytes but much worse than that with a M S S of 1458 bytes. It seems that a larger M S S can improve the throughput performance of a l l T C P implementations over a whole range of the P L R in the BSI I environment. However, another limitation of T C P throughput performance is the maximum transfer unit ( M T U ) in byte of the interfaces. When the IP layer receives an IP datagram, It not only determines which local interface the datagram is being sent on but also queries that interface to obtain its M T U . IP compares the datagram size with the M T U and perform fragmentation i f the datagram size is bigger than the M T U . When an IP datagram is fragmented, it is not reassembled until it reaches its final IP layer at the destination. In this case the M T U in a typical Ethernet L A N is 1516 bytes. Beyond this limitation of the M T U , the transit ethernet frame has a relatively high probability of Chapter 4 Analysis of TCP in a Satellite Environment 62 framentation. Subtracting the overheads of the ethernet frame and IP datagram, the M S S , in this case, is 1458 bytes. A M S S of 1458 bytes seems to give higher throughput for all T C P implementations. With a M S S of 1458 bytes, T C P Vegas has 1.3 times better throughput than T C P Reno and nearly six times better throughput than T C P R F C at a relatively low P L R . A t a relatively high P L R , T C P Vegas has approximately two times better throughput than T C P Reno and five times better throughput than T C P R F C . Overall, T C P Vegas has much better throughput performance than the other T C P implantations for different M S S s over the whole range of P L R s in the satellite channel. •S V Segment Size = 512 Bytes Segment Size = 1024 Bytes _ Segment Size = 1458 Bytes Segment Size = 1460 Bytes O . O I 0 . 0 2 0 . 0 3 0 . 0 4 O . O S 0 . 0 6 0 . 0 7 P a c k e t L o s s R a t © i n t h e S a t e l l i t e L i n k s Figure 4.19 Throughput of TCP R F C vs. the PLR in the satellite links for different MSSs Segment Size = 512 Bytes . . _ Segment Size = 1024 Bytes _ Segment Size = 1458 Bytes Segment Size = 1460 Bytes O . O I 0 . 0 2 0 . 0 3 0 . 0 4 0 . 0 5 0 . 0 6 0 . 0 7 I P a c k e t L o s s R a t e i n t h e S a t e l l i t e L i n k s Figure 4.20 Throughput of T C P Reno vs. the PLR in the satellite links for different MSSs Chapter 4 Analysis of TCP in a Satellite Environment 63 1 4 0 K>\ Segment Size = 512 Bytes Segment Size = 1024 Bytes Segment Size = 1458 Bytes Segment Size = 1460 Bytes o 0.01 0 . 0 2 0 . 0 3 0 . 0 4 0 . 0 5 P a c k e t L o s s R a t e i n t h e S a t e l l i t e L i n k s 0 . 0 6 O . O S Figure 4.21 Throughput of T C P Vegas vs. the PLR in the satellite links for different MSSs 4.7 Optiminzation and Discussion In order to improve throughput performance of different T C P implementations for a range of the P L R s in the satellite links, this section provides a detailed evaluation for jointly optimizing the T C P parameters including the M T W S , the M A D , and the M S S . From the simulation results in section 4.6, an upper bound and a lower bound of the T C P parameters are obtained and suggested for T C P R F C , T C P Reno and T C P Vegas under the assumptions in section 4.3 for different P L R s . In this section, all other simulation parameters are fixed, but the above T C P parameters are varied under different conditions. The R T T coefficient deviation is assumed to be 4 and the thresholds ( a and (3) in T C P Vegas are assumed to be 1 and 3. g is assumed to be 0.05. b is assumed to be 0.01. A n average E T E delay of 0.3 sec is assumed in the first scenario and a P L R of 0.001 is assumed in the second scenario. Since these T C P parameters do not have great effects on the throughput of T C P R F C , the throughput values of T C P R F C for different T C P parameter settings are too close to be statisti-cally different. A bigger M T W S results in increasing packet losses in the Internet due to network Chapter 4 Analysis of TCP in a Satellite Environment 64 congestion. A smaller M A D and a larger segment size seem to improve the throughput perfor-mance for T C P R F C . Therefore, over the whole range of the P L R s , the optimal values of the M T W S , the M A D , and the M S S are suggested to be 16 Kbytes, 0.1 msec, and 1458 bytes, respec-tively in this H W W network. Figure 4.22-4.24 show the throughput of T C P Reno as a function of the P L R in the satellite channel for different M S S s with different T C P parameter settings. From those figures, for different M T W S s and different M A D s , T C P Reno with a M S S of 1458 bytes seems to have much better throughput than that with the other M S S s over the whole range of P L R s in the satellite channel. T C P Reno with a M S S of 512 bytes seems to have much lower throughput than that with the other M S S s . For different M T W S s and different M A D s , T C P Reno with a M S S of 1458 has 2.5 times better throughput than that with a M S S of 512 bytes and has approximately 1.5 times better throughput than that with both M S S s of 896 bytes and 1024 bytes. However, T C P Reno with a M S S of 1024 bytes is just slightly better throughput than that with a M S S of 896 bytes. Q\ 1 1 I 1 1 1 1 1 1 1 1 1 Q! 1 I I I I I I ' ' I L_ 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 Packet Loss Rate in the Satellite Links Packet Loss Rate in the Satellite Links Figure 4.22 Throughput of T C P Reno vs. PLR in the satellite links for different MSSs (MTWS = 64 Kytes) Chapter 4 Analysis of TCP in a Satellite Environment 65 80 i i l l l l l l l i i 80 i i i i i i i i i i i M S S = 512 Bytes M S S = 512 Bytes 70 M S S = 896 Bytes 70 • , . M S S = 896 Bytes t. • M S S = 1024 Bytes \\ '. M S S = 1024 Bytes U ': " M S S = 1458 Bytes I ' l \ - - - . M S S = 1458 Bytes 60 ' y \ 60 \V '•• « "a ID I \ \ '•, f 50 • 1 1 \ 5. \ \ - • n \ V \ . M A D = 1 msec .0 * \ \ ''••. M A D = 2 msec 1 4 0 • \ \ \ | 4 0 " ^ \ \ . v V 0) x \ \ 0 V V 0 30 30 x \ \ v v 20 - \ 20 ^ V \ \ 10 10 1 1 1 1 1 1 1 1 1 1 1 0 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 Packet Loss Rate in the Satele Lmks Packet Loss Rale in the Satellite Links Figure 4.23 Throughput of TCP Reno vs. PLR in the satellite links for different MSS (MTWS = 80 Kbytes) 80 i i i i i i i I i > i 80 ' . 1 1 1 1 1 I I I 1 ! I 1 •• M S S = 512 Bytes 1 '•. M S S = 512 Bytes 70 ', • M S S = 896 Bytes \ ' l '•. M S S = 1024 Bytes \'l • " M S S = 1458 Bytes 70 - 1 • M S S = 896 Bytes \ \ M S S = 1024 Bytes 1 • M S S = 1458 Bytes 60 • 1 N '• 60 1 \' (j "50 ' \ x '• ' 1 \ \ '•. in 8 50 "1 I | 1 \ ^ ' K ' \ \ ' V\ 140 ' ' \ \ 1 4 0 \\ \ \ UI 3 i \ \ M A D = 1 msec £. Ol \ \ \ M A D = 2 msec O i \ \ O \ y H30 \ \\ \ V\ \ V x V H30 \ \\ \ y \ \'\ \ \'\ 20 • \ V 20 ^ \\ \ y - -10 10 ~ ~ ~ ^ . ^ 0 1 1 ( 1 p 1 1 1 1 1 t 0 i t i i i i i i i i i 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 Packet Loss Rate in the Satellite Links 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 Packet Loss Rate in the Satellite Links Figure 4.24 Throughput of T C P Reno vs. PLR in the satellite links for different MSS (MTWS = 128 Kbytes) Figure 4.25 shows the throughput of T C P Reno with a M S S of 1458 bytes as a function of the P L R for different M T W S s and different M A D s . A t low P L R s (less than 0.03 in this case), Chapter 4 Analysis of TCP in a Satellite Environment 66 T C P Reno with a M A D of 2 msec has higher throughput than that with a shorter M A D for differ-ent M T W S s . However, T C P Reno with a M A D of 1 msec has slightly better throughput than that with a longer M A D for different M T W S s at a relatively high P L R in the satellite link. 90 I 1 1 1 1 1 1 r 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 Packet Loss Rate in the Satellite Link Figure 4.25 Throughput of TCP Reno with a segment size of 1458 bytes vs. the PLR in the satellite link A t a low P L R (< 0.03 in this case), T C P Reno with a M T W S of 80 Kbytes has much better throughput than that with other M T W S s . In the range of the P L R between 0.03 to 0.07, the throughput of T C P Reno with a M T W S of 80 Kbytes is slightly better than that with the other M T W S s . Beyond a P L R of 0.07, the throughput of T C P Reno with different M T W S s and differ-ent M A D is too close to be statistical significance. Figure 4.25 proves that the upper and lower bounds of the M T W S and the M A D do not have significant effects on the throughput performance of T C P Reno. In general, a M T W S of 80 Kbytes, a M A D of 2 msec, and a M S S of 1458 bytes give a reasonably high throughput for T C P Reno over a whole range of P L R s in the satellite channel for this H W W environment. Chapter 4 Analysis of TCP in a Satellite Environment 67 Figure 4.26-4.28 show the throughput of T C P Vegas as a function of the P L R in the satellite links for different M S S s under different T C P parameters settings. For different M A D s , T C P Vegas with a 1458 byte M S S seems to have much better throughput then that with the other M T W S s (128 Kbytes and 80 Kbytes) but slightly better throughput than that with a 64 Kbyte M T W S . However, T C P Vegas with a 512 byte M S S has much lower throughput performance than that with the other MSSs . For different M A D s , with , T C P Vegas with a 1458 byte M S S and a 128 Kbyte M T W S or a 80 Kbyte M T W S has roughly double better throughput than that with a 512 byte M S S , 1.5 times better throughput than that with a 896 byte segment size, and slightly better throughput that with a 1024 byte segment size. For different M A D s , the M S S only has a slight effect on the throughput of T C P Vegas with a 64 Kbyte M T W S over the whole range of P L R s in the satellite channel. no 100 X 70h e 50 - l 1 1 1 1 1 1 1 1 1 r M S S = 512 Bytes M S S = 896 Bytes M S S = 1024 Bytes *\ . . . . M S S = 1458 Bytes M A D = 1 msec _ i i i i i i_ 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 Packet Loss Rate in the Satellite Links -t 1 1 1 1 1 r M S S = 512 Bytes M S S = 896 Bytes M S S = 1024 Bytes ( \ - - - - M S S = 1458 Bytes M A D - - 2 msec _! I I 1_ 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 Packet Loss Rate in the Sateie Links Figure 4.26 Throughput of T C P Vegas vs. PLR in the satellite links for different MSS (MTWS = 64 Kbytes) Chapter 4 Analysis of TCP in a Satellite Environment 68 110 110 i .i . i i i — i 1 1 T— "' • 1 1 100 90 80 1 \ \ \ ' l \ \ \ M S S = 512 Bytes 1 \ \ M S S = 896 Bytes - 1 \ \ - . M S S = 1024 Bytes \ \ \ \ - - - - M S S = 1458 Bytes \ \ \ 100 90 80 \ U • \ \ \ \ •-V \ \ \ \ '•• . . \ \ '\ • \ \ '\ M S S = 512 Bytes M S S = 896 Bytes M S S = 1024 Bytes M S S = 1458 Bytes S 70 vs 7> ID £ 60 \ \ s * \ \ v \ '\ • • ^ \ \ 8 70 I 0 %• 60 •\ \ \ '•. \ \ \ ' . \ \ \ ' . v \ \ '•• ' \ '•• • jghput (Kt \ \ \ M A D - 1 msec v \ \ v V N jghput (Kt o I • \ \ w M A D = 2 msec Throi 0 jE 40 • \ \ \ \ \ 30 30 • \ \ ~ - . 20 20 • ~ ~ ~ ^ 10 10 0 1 1 1 1 1 p i i i i i 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 Packet Loss Rate in the Satellite Links 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 Packet Loss Rate in the Satellite Links Figure 4.28 Throughput of T C P Vegas vs.PLR in the satellite links for different MSS ( M T W S = 128 Kbytes) Figure 4.29 shows the throughput of TCP Vegas with a 1458 bytes MSS as a function of the PLR in satellite link for different MTWSs and different MADs. The throughput of TCP Vegas Chapter 4 Analysis of TCP in a Satellite Environment 69 with a M A D of 10 msec, overall, is better than that with other M A D s for different M T W S s over the whole range of the P L R s . A t a low P L R (less than 0.07 in this case), T C P Vegas with a 128 Kbyte M T W S has much better throughput than that with other M T W S s , whereas T C P Vegas with a 64 Kbyte M T W S has much worse throughput than that with other M T W S s over the whole range of P L R s in the satellite channel. Beyond a P L R of 0.07, T C P Vegas with a 128 Kbyte M T W S and a M A D of 10 msec is still slightly better than that with other M T W S s 120 100 80 60 40 20 M T W S = M T W S = M T W S = M T W S = M T W S = M T W S = M T W S = M T W S = 64 Kbytes, M A D = 10 msec 80 Kbytes, M A D = 10 msec 128 Kbytes, M A D = 10 msec 64 Kbytes, M A D = 20 msec 80 Kbytes, M A D = 20 msec 128 Kbytes, M A D = 20 msec 80 Kbytes, M A D = 5 msec 128 Kbytes, M A D = 5 msec X " . 0.01 0.02 0.03 0.04 0.05 0.06 Packet Loss Rate in the Satellite Link 0.07 0.08 Figure 4.29 Throughput of T C P Reno with a segment size of 1458 bytes vs. the PLR in the satellite links Compared with figure 4.25, figure 4.29 shows that the upper and lower bounds of the M T W S and the M A D have greater effect on throughput performance of T C P Vegas than on that of T C P Reno. In general, T C P Vegas with a 128 Kbyte M T W S , a 10 millsecond M A D , and a 1458 byte M S S , overall, has a reasonably high throughput performance over a whole range of P L R s in the satellite channel. Finally, the jointly optimal values of the chosen T C P parameters for T C P Vegas, T C P Reno and T C P R F C with a 0.3 second average E T E delay for different P L R s in the satellite channel are summarized in table 4.4. Chapter 4 Analysis of TCP in a Satellite Environment 70 Table 4.4 The suggested optimal TCP paramenters for different T C P implementations T C P Parameters TCP Vegas TCP Reno T C P R F C M T W S 128 Kbytes 80 Kbytes 16 Kbytes M A D 10 msec 2 msec 0.5 msec MSS 1458 Bytes 1458 Bytes 1458 Bytes The throughput performance of the T C P implementations with their optimal T C P parame-ters as a function of the P L R and the average E T E delay are presented in figure 4.30 and 4.31, respectively. Figure 4.30 shows that the throughput of al l T C P implementations with their optimal T C P parameters is st i l l decreased gradually when the P L R in the satellite channel is increased. The state model nearly has no effect on the throughput performance of T C P R F C , a slight effect on the throughput of T C P Reno, a great effect on the throughput of T C P Vegas. Compared with figure 4.3, figure 4.30 shows that the throughput performance of the different T C P implementations is significantly improved after jointly optimizing the T C P parameters for each T C P implementations over a whole range of P L R s . T C P Vegas with its optimal parameters has an approximately 34% to 100% better throughput than that without optiminzing its T C P parameters when the P L R varies from 0.001 to 0.06. T C P Reno with its optimal T C P parameters has an approximately 76% to 500% better throughput than that without optiminzing its T C P parameters when the P L R varies from 0.001 to 0.06. T C P R F C with its optimal T C P parameters, in general, has a approximately 300% better throughput than that without optiminzing its T C P parameters over the whole range of P L R s . Table 4.5 shows the throughput comparison of T C P Vegas, T C P Reno and T C P R F C with their optimal parameters under a 0.3 second average E T E delay condition. From table 4.5, T C P Vegas has 23% better throughput than T C P Reno and more than 700% better throughput than T C P R F C under a clear sky condition. A t a P L R of 0.06, T C P Vegas still has more than 200% better throughput than T C P Reno and around 500% better throughput than T C P R F C . Chapter 4 Analysis of TCP in a Satellite Environment 71 100 Single State Model Two State Model 0 . 0 0 5 0 . 0 1 0 . 0 1 5 0 . 0 2 0 . 0 2 5 0 . 0 3 0 . 0 3 5 0 . 0 4 0 . 0 4 5 0 . 0 5 0 . 0 5 5 P a c k e t L o s s R a t e in t h e S a t e l l i t e L i n k Figure 4.30 Throughput of different T C P implementations vs. the PLR in the satellite channel Table 4.5 Throughput in Kbytes/sec of TCP versions with the E T E delay of 0.35 sec T C P Vegas T C P Reno T C P R F C Throughput under a clear sky condition 170.31 138.56 23.06 Throughput ratio 7.386 6.008 1.0 Throughput at PLR of 0.02 63.61 25.66 10.31 Throughput ratio 6.170 2.489 1.0 Throughput at PLR of 0.04 37.42 14.31 7.3 Throughput ratio 5.126 1.960 1.0 Throughput at PLR of 0.06 26.16 10.26 5.65 Throughput ratio 4.630 1.816 1.0 Figure 4.31 shows the E T E delay has greater effect on T C P Vegas and T C P Reno than on T C P R F C ; whereas the E T E delay has much less effects on the throughput of T C P R F C than on the others. When the E T E delay increases, the throughput of all T C P implememations is linearly decreased. Compared with figure 4.5, figure 4.31 show that T C P Reno with its optimal T C P parameters has significantly improvement in its throughput than the other T C P implementations. Chapter 4 Analysis of TCP in a Satellite Environment 72 Table 4.6 shows the throughput comparison of T C P Vegas, T C P Reno and T C P R F C with their optimal parameters at a 0.001 P L R in the satellite channel. After jointly optiminzing the key T C P parameters for the different T C P implementations, table 4.6 shows that T C P Vegas still has nearly 40% better throughput than T C P Reno and more than 700% better throughput than T C P R F C over the whole range of average E T E delays. However, all T C P implementations perform badly under a large E T E delay condition. 1 6 0 1 4 0 1 2 0 J» 1 0 0 CD •R ~ 8 0 ZD CX. O) 2 6 0 4 0 2 0 I 1 1 1 1 _ ~ -V e g a s --R F C -•"• - x X t 1 1 1 1 1 1 1 0 . 4 5 0 .5 0 . 5 5 0 .6 0 . 6 5 0 .7 0 . 7 5 0 .8 0 . 8 5 0 .9 A v e r a g e end—to—end d e l a y Figure 4.31 Throughput of different TCP implementations vs. the average E T E delay at a 0.001PLR Table 4.6 Throughput in Kbytes/sec of different TCP versions at a PLR of 0.001 T C P Vegas T C P Reno T C P R F C Throughput at ete delay of 0.35 147.82 105.52 18.57 Throughput ratio 7.96 5.68 1.0 Throughput at ete delay of 0.65 sec 102.23 69.13 8.87 Throughput ratio 11.53 7.79 1.0 Throughput at ete delay of 0.85 sec 81.69 53.45 6.71 Throughput ratio 12.17 7.97 1.0 Chapter 5 Analysis of TCP in a Mobile Environment In this chapter, the performance of two T C P implementations (TCP Vegas and T C P Reno) in an environment of mobile radio channels interworking with the Internet (MRCII) is studied and compared. Optimization of some key T C P parameters for the T C P implementations in this M R C I I environment under different conditions is presented. 5.1 Cellular Mobile Communications A schematic topology of a mobile computing environment consisting of wireless data networks interworking with the Internet is shown in Figure 5.1. The wide geographic area is divided into radio cells. Each cell has a base station, providing the radio interface with mobile hosts within its coverage area. Each base station is connected either to a switching office such as a mobile switching centre or directly to a gateway. Gateways may connect the remote mobile hosts via the Internet to fixed hosts attached to L A N s . Figure 5.1 Interconnection between a mobile host and Internet fixed hosts 73 Chapter 5 Analysis of TCP in a Mobile Environment 74 The mobile switching center has capability to separate the voice and data traffic, and data packets are processed and forwarded over the Internet. For example, in the C D P D system, this function is performed by an intermediate system called the Mobi le Data Intermediate System. The T C P data segment transmission over the Internet is described in section 5.1. A gateway is used to connect the mobile data network with the Internet. When the gateway receives a packet from the Internet for a mobile host which it currently controls, it simply forwards the packet to an appropriate base station through which the mobile host can be reached. When expected packets are received in the mobile host, the corresponding acknowledgment packets are returned immedi-ately from the destination mobile host to the source node. F r o m the simulation results in chapter 4, T C P R F C performs poorly in a high P L R environment. Therefore, only T C P Vegas and T C P Reno are considered in the following discus-sion. Because of the variable propagation changes environment and limited maximum data rate in mobile data networks, optimization of a number of T C P dependent parameters for both T C P implementations are necessary to improve T C P throughput. Aspects of communications related to the data link and the physical layer are beyond the scope of this thesis; therefore, no data link layer error recovery is assumed. 5.2 Simulation Model Figure 5.2 shows the simulation model for a typical IP network incorporating with the mobile data networking environment. The simulation model includes a number of nodes and is divided into a mobile radio section and a Internet section and interconnected by a gateway. The mobile radio section models the characteristics of the mobile channel described in section 3.2. The Internet section models the characteristics of the Internet traffic described in section 3.3. In the simulation model, data packets are sent from a fixed source node interconnected with a L A N Chapter 5 Analysis of TCP in a Mobile Environment 75 to a destination mobile host, and the corresponding A C K packets are returned immediately from the destination mobile host to the source node. Passing through the Internet section and the mobile radio section, the data and A C K packets are delayed and may be lost due to congestion, transmission errors in the media, or handoffs. Figure 5.2 The mobile data network simulation model 5.3 Simulation Assumptions and Parameters The simulation model is implemented in O P N E T based on assumptions described in section 4.3. The fixed simulation parameters of the Internet section for all the simulations in this chapter are summarized in table 4.1. However, based on the simulation results of [7], the big window option need not be used in T C P connections with low bandwidth-delay products, as is the case here; therefore, the size of the T C P header is 20 bytes. The maximum window size in the receiver is assumed to be the same as that in the transmitter. The B E R in the bad state within the normal state is much higher than that in the good state. The maximum data rate in the mobile data channel is assumed to be 19.2 Kbps. In order to maximize T C P performance over the mobile data Chapter 5 Analysis of TCP in a Mobile Environment 76 network interconnected with the Internet, the effects of the mobile speed, the P L R in the bad state within the normal state, the average time between handoffs (1/a), and the average time in performing handoff (1/b) on T C P throughput need to be analyzed. Moreover, key T C P parame-ters including the M T W S , the M A D , and the M S S need to be jointly optimized under different conditions. 5.4 Impact of Mobile Fading on TCP Throughput This section presents the effects of the mobile speed and the P L R due to unreliable mobile radio links on the throughput performance of T C P Reno and T C P Vegas. Because of the signal-fading phenomenon, the average duration of fades (1/d) is inversely proportional to the mobile speed,.and the S T P from a good state to a bad state within the normal state ( l -e" c t ) is directly proportional to the mobile speed (section 3.2), where 1/c is the average time between fades. When the mobile unit moves fast, the rate of the signal fluctuation is fast, but 1/d is short, thus causing a relatively high S T P between the good state and the bad state even when there is no handoff involved. As a result, the loss probability for every transmitted packet is relatively high. In this simulation, all the other key T C P parameters are fixed with values as suggested in [7] [8], but the mobile speed and the P L R in the bad state within the normal state is varied. The mobile unit is assumed to be moving only within a cell so that no handoff process is involved. The M S S is assumed to be 512 bytes. The M T W S is assumed to be 8 Kbytes and the M A D is assumed to be 1 msec (typical use in Ethernet L A N s ) . The R T T deviation coefficient is assumed to be 4. The thresholds ( a and (3) in T C P Vegas are assumed to be 1 and 3. Figure 5.3 shows the average P L R in the mobile radio channel as a function of the mobile speed for different P L R s in the bad state within the normal state. Figure 5.4 shows the throughput of T C P Reno and T C P Vegas as a function of the mobile speed for different P L R s in the bad state within the normal Chapter 5 Analysis of TCP in a Mobile Environment state, respectively. 77 In figure 5.3, the average P L R in the mobile channel increases when the mobile speed is gradually increased because the STPs between the good state and bad state within the normal state are increased accordingly. In figure 5.4, the throughput of T C P Reno and T C P Vegas is gradually decreased when the mobile speed is increased for different P L R s in the bad state. B y increasing the mobile speed, the rate of signal fluctuation is increased, thus causing a relatively high STP from the good state to the bad state within the normal state. As a result, the P L R over the mobile channel is relatively high even with a shorter 1/d. The mobile speed seems to have greater effects on the throughput of T C P Reno than on that of T C P Vegas. Over the whole range of mobile speeds, T C P Vegas seems to have slightly better throughput than T C P Reno for different P L R s in the bad state within the normal state. O . O I 6 , , , | | , q 3s - *•* . • S o.ois - — ~~ M f • g g 0 . 0 1 4 — c : - § 2 ^ 0 . 0 1 3 " 0 a> PLR in the bad state = 0.05 ffi 0 . 0 1 2 2 PI ,R in the had state = 0.02 !£ ca O . O I 1 Average p ± , I , 1 1 1 1 1 *• 1 ' 2 0 3 0 4 0 S O 6 0 7 0 S O 9 0 1 0 O 1 1 0 M o b i l e S p e e d ( K m / h ) Figure 5.3 Average PLR in the mobile radio channel vs. mobile speed Figure 5.5 shows a more detailed picture of the effects of the P L R in the bad state within the normal state on the throughput of T C P Reno and T C P Vegas at different mobile speeds. At a mobile speed of 45 km/h, T C P Vegas has approximately 0.5% to 4% better throughput than T C P Chapter 5 Analysis of TCP in a Mobile Environment 78 Reno when the P L R in the bad state varies from 0.001 to 0.35. However, at a mobile speed of 75 km/h, T C P Vegas has approximately 0.7% to 10% better throughput than T C P Reno when the P L R in the bad state varies from 0.001 to 0.35. Therefore, while too low a mobile speed may increase the 1/d but decrease the STPs between the two states, and too high a mobile speed may increase the rate of signal fluctuation. From figure 5.5, T C P Vegas seems to have a better response to the increased rate of signal fluctuation because of its faster error recovery and better congestion avoidance mechanisms. 2600 2200; o £2000 £ 1(800 1200 1000 P L R in the bad state = 0.02 P L R in the bad state = 0.05 P L R in the bad state = 0.2 P L R in the bad state = 0.35 120001 50 60 70 80 90 1 00 110 1 20 Mobile Speed (Km/h) 12001 1000 P L R in the bad state = 0.02 P L R in the bad state = 0.05 P L R in the bad state = 0.2 P L R in the bad state = 0.35 T C P Vegas 40 50 60 70 80 90 100 110 120 Mobile Speed (Km/h) Figure 5.4 TCP Throughput vs. mobile speed for different PLRs in the bad state Figure 5.6 shows the congestion window response in both T C P Reno and T C P Vegas as a function time for different mobile speeds at a P L R of 0.05 in the bad state within the normal state. When the mobile unit moves slowly in a relatively high P L R environment, 1/d is relatively long, thus causing several packets to be corrupted per each full transmit window size. As a result, not only the slow-start mechanism but also the exponential back-off algorithm are relatively active, causing unnecessarily long communication pauses. When the mobile unit moves fast in a Chapter 5 Analysis of TCP in a Mobile Environment 79 relatively high fading environment, the rate of the congestion window fluctuation is increased, thus the congestion window does not have enough time to expand before the next mobile radio channel interruption occurs. 3 Sr 1 6 0 0 T C P Reno. Mobile speed = 45 Km/h T C P Reno. Mobile speed = 75 Km/h T C P Vegas. Mobile speed = 45 Km/h T C P Vegas, Mobile speed = 75 Km/h O . O S O . I 0 . 1 S 0 . 2 0 . 2 5 0 . 3 P a c k e t L o s s R a t e i n t h e b a d s t a t e w i t h i n t h e N o n — H a n d o f f s t a t e Figure 5.5 TCP throughput vs. PLR in the bad state for different mobile speeds >> CQ Mobile speed = 45 Km/hr TCP Vegas Mobile speed TCP Reno = 45 Km/hr 1.2S l . S 1.7S 2 t i n * (itc) (xlOOO) 1.75 2 (sec) (xlOOO) Backoff Period CQ o •a Mobile speed = 75 Km/hr TCP Vegas 1.2S 1.5 1.75 2 time (fee) (XlOOO) 1.75 2 ec) (xlOOO) Figure 5.6 The congestion window response in TCP for different mobile speeds Chapter 5 Analysis of TCP in a Mobile Environment go Because of the longer 1/d, the exponential back-off reaction in T C P Reno at a mobile speed of 45 km/hr seems more active (see backoff period indicated with an arrow in figure 5.6) than that at 75 Km/h , degrading T C P throughput. At 45 Km/h, T C P Vegas has less frequent drops (reset to one M S S ) of the congestion window than that at 75 K m / h . Because T C P Vegas has a better congestion avoidance and a faster error recovery mechanism, the rate of the congestion window fluctuation in T C P Vegas is much less than that in T C P Reno. The size of the congestion window in T C P Reno is limited by not only mobile radio channel noise but also congestion. 5.5 Impact of Handoff on TCP Throughput This section presents the effects of the handoffs due to cell crossings on the throughput performance of T C P Reno and T C P Vegas. The 1/a is directly proportional to the size of the radio cells and inversely proportional to the mobile speed. The 1/b is directly proportional to the area of the handoff regions and inversely proportional to the mobile speed. A l l the T C P parameters are fixed to the same values as in section 5.4 but only the 1/a and the 1/b are varied by varying the radius of the radio cells (from 0.5 K m to 5 Km) and the handoff regions for different mobile speeds. For every 1/a sec on average, when the mobile host moves from one cell to another cell, the wireless channel is assumed to be unavailable (in the handoff state) for an average of 1/b sec. A n average P L R in the bad state within the normal state is assumed to be 0.05. Figures 5.7, 5.8, and 5.9 show the throughput performance of T C P Reno and T C P Vegas as a function of 1/a for a mobile speed of 45 Km/h , 75 Km/h , and 120 Km/h , respectively, under different 1/b values, where the handoff region = 50 m, 100 m and 200 m are considered. They show that 1/a and 1/b have a great influence on the throughput of both T C P Reno and T C P Vegas for different mobile speeds. A s expected, the throughput of both T C P implementations is decreased as 1/b is increased and increased as 1/a is increased at different mobile speeds. A Chapter 5 Analysis of TCP in a Mobile Environment 81 shorter 1/a at faster mobili ty rate causes a larger degradation of the throughput of both T C P implementations. A longer handoff delay has a higher P L R in the mobile radio link. A t different mobile speeds, T C P Vegas still has better throughput than T C P Reno over the whole range of 1/a for different 1/b values. For different handoff regions and different sizes of the radio cells, the throughput of T C P Reno and T C P Vegas decreases approximately 2% when the mobile speed is increased from 45 K m / h to 75 Km/h and approximately 6% when the mobile speed is increased from 45 K m / h to 120 Km/h . m 1 6 0 0 TCP Reno TCP Vegas - O — l/b = 2.8sec —t— l/b = 5.8 sec -x— l/b = 8.6 sec 1 O O 1 2 0 1 4 0 1 6 0 1 8 0 A v e r a g e T i m e b e t w e e n h a n d o f f s ( s e c ) Figure 5.7 TCP throughput vs l/a for different l/b values at a mobile speed of 45 Km/h 8 S O l O O 1 2 0 1 4 0 1 6 0 A v e r a g e T i m e b e t w e e n h a n d o f f s ( s e c ) Figure 5.8 T C P throughput vs l/a for different l/b values at a mobile speed of 75 Km/h Chapter 5 Analysis of TCP in a Mobile Environment 82 TCP Reno TCP Vegas 1/b = 1.2 sec 1/b = 2.4 sec 1/b = 4.8 sec 2 0 4 0 6 0 S O 1 0 0 1 2 0 1 4 0 1 6 0 1 B O 2 0 0 A v e r a g e T i m e b e t w e e n h a n d o f f s ( s e c ) Figure 5.9 TCP throughput vs 1/a for different 1/b values at a mobile speed of 120 Km/h Figure 5.10 shows a detailed graph of the effects of 1/a on the throughput performance of T C P Reno and T C P Vegas as a function of 1/a for different 1/b values at a mobile speed of 75 K m / h. It shows that the throughput of both T C P implementations is seriously affected by 1/a. As 1/a is reduced, for different 1/b values, the throughput of both T C P implementations decreases because there are more and more retransmissions. When 1/a is further reduced, the recovery of the T C P connection from a channel interruption is not fast enough to expand the current conges-tion window size before the next channel interruption occurs. These channel interruptions violate most of the conditions required for reliable T C P operation. As a result, an incorrect T C P time-out for every segment is estimated and causes the serious degradation of T C P throughput. However, for a larger 1/a, the duration of the channel availability is sufficiently long to open up the congestion window size before the next interruption occurs. Therefore, there is sufficient bandwidth not only to retransmit all the lost segments but also to send more segments into the network. T C P Vegas, overall, has approximately 25% to 14% better throughput than T C P Reno for a smaller 1/b and has approximately 12% to 5% better throughput than T C P Reno for a larger 1/b when 1/a varies from 35 sec to 180 sec. Chapter 5 Analysis of TCP in a Mobile Environment 83 S O T O O 1 2 0 1 4 0 A v e r a g e t i m e b e t w e e n h a n d o f f s ( s e c ) Figure 5.10 TCP throughput vs. l/a for different l/b values at mobile speed of 75 Km/h Figure 5.11 plots the sequence numbers of the received T C P segments as a function of time for l/b = 3.5 sec and l /a = 70 sec at a mobile speed of 75 K m / h to present a more detailed graph of the behavior of the end-to-end connection for T C P Vegas and T C P Reno. T C P Vegas maintains a higher and more consistent throughput than T C P Reno. Because T C P Reno unneces-sarily invokes the congestion control procedures more often than T C P Vegas for the duration of the connection, the extensive communication pauses significantly degrade T C P throughput. This phenomenon appears as the flat and empty regions of the curve. Figure 5.12 plots the congestion window response in T C P Reno and T C P Vegas as a function of time for l/b = 3.5 sec with different 1/a values at a mobile speed of 75 K m / h to present the behavior of the congestion window response in both T C P implementations for different 1/a values. T C P Reno, as expect, has more frequent drops of the congestion window than T C P Vegas for different 1/a values. When 1/a = 70 sec, both T C P implementations invoke the slow-start mechanism more often and have longer communication pauses than both T C P implementations when 1/a = 200 sec, thus degrading the T C P throughput performances. The flat and empty regions of the curve represent the communication pauses. Chapter 5 Analysis of TCP in a Mobile Environment 84 Seepience Bomber (xle+06) Sequence Homier (xle+06) 3.S j 1 3.5 time (sec) (xlOOO) time (sec) (xlOOO) Figure 5.11 Sequence numbers for TCP segments transferred to mobile host over the mobile channel Tim. (sac) Tlm« (sec) Figure 5.12 Congestion window response in TCP over mobile channel at a mobile speed of 75 Km/h Chapter 5 Analysis of TCP in a Mobile Environment 85 5.6 Impact of TCP Parameters on TCP Throughput This section analyzes the effects of some key T C P parameters including the M T W S , the M A D , and the M S S on the throughput performance of both T C P Reno and T C P Vegas over mobile radio channels with severe fading and frequent handoffs, internetworking with the Internet. In this section, the thresholds ( a and (3 ) in T C P Vegas are assumed to be 1 and 3 [26], and the R T T deviation coefficient is assumed to be 4 [17]. 5.6.1 Effects of Maximum Transmit Window Size Many pervious studies have showed that the M T W S is a key T C P parameter which has a strong influence on T C P throughput in the mobile fading environment [6][7][8][10]. In this subsection, the effects of the M T W S on the throughput performance of both T C P implementa-tions are analyzed and compared. The optimal M T W S s for both T C P implementations to maximize throughput under different conditions are presented. A l l other network components and T C P parameters are fixed for performance evaluation by only varying the M T W S . The M S S is assumed to be 512 bytes as suggested in [7]. The M A D is assumed to be 1 msec. In this subsection, two scenarios are considered. In the first scenario, the radius of the radio cells and the area of the handoff region are fixed at 2 K m (a lower bound of radio marco cells) and 100 m, respectively, but the speed of the mobile unit is varied. In the second scenario, the mobile speed is fixed at a 75 Km/h , but 1/a and l/b are varied. The P L R in the bad state within the normal state is assumed to be 0.02 for both scenarios. Figure 5.13 shows the throughput of both T C P implementations in the first scenario as a function of the M T W S for different mobile speeds. Figure 5.14 and figure 5.15 show the throughput of both T C P implementations in the second scenario as a function of the M T W S for different 1/a and l/b values, respectively. Chapter 5 Analysis of TCP in a Mobile Environment 86 Figure 5.13 shows that the M T W S seriously affects the throughput of both T C P implementations at different mobile speeds. The throughput of both T C P implementations is increased when the M T W S is increased. Beyond a certain M T W S , their throughput gradually decreases. At different mobile speeds, both T C P implementations with a smaller M T W S (< 8 Kbytes) have better throughput than that with a larger M T W S . However, too small a M T W S has greater effects on the throughput of T C P Vegas than on that of T C P Reno especially when the mobile unit moves significantly fast. Too small a M T W S limits T C P sending more data into the network when the channel staying in the normal state, thus seriously degrading T C P throughput. A t a fast relatively mobile speed, both T C P implementations prefer to use a smaller M T W S because of the increasing rate of signal fluctuation. On the other hand, they prefer to use a larger M T W S at a relatively slow mobile speed in order to send more data segments when the channel staying in the normal state. A t different mobile speeds, T C P Reno prefers to use a smaller M T W S than T C P Vegas because of the limitation of its congestion avoidance mechanism. 1 0 1 5 2 0 M a x i m u m T r a n s m i t W i n d o w S i z e ( K B y t e s ) Figure 5.13 TCP throughput vs. MTWS for different mobile speeds Figures 5.14 and 5.15 show that the M T W S has great effects on the throughput of both Chapter 5 Analysis of TCP in a Mobile Environment 87 T C P implementations for different 1/a and l /b values. Figure 5.14 shows that both T C P implementations prefer to use a bigger M T W S for a longer 1/a and a smaller transmit window size for a shorter 1/a. It is because a bigger M T W S allows more data packet sending into the network during the time between handoffs. However, too large a M T W S requires much longer recovery time from the channel interruptions for the source host to go back to the nominal throughput level. With a bigger M T W S , both T C P implementations have difficulty recovering from multiple losses within each full sending window by using their error recovery mechanisms; therefore, most of the lost packets are recovered by the coarse-grain RTO mechanism. A smaller M T W S seems to give much better T C P throughput performance in this M R C I I environment. Figure 5.15 shows that both T C P implementations prefer to use a smaller M T W S for both l/b considered values since a bigger transmit window size creates more packet losses during the time performing handoffs. Due to their congestion avoidance mechanisms, a bigger M T W S causes a long communication pause. It shows that a 4 Kbyte M T W S seems to give much better throughput for both T C P implementations for different l/b values at a mobile speed of 75 Km/h . 2300 22SO \ -TCPReno , 1/a = 70 sec T C P Reno, 1/a = 140 sec T C P Reno, 1/a = 210 sec T C P Vegas, 1/a = 70 sec 1900 1850 1 800 20 30 40 50 Maximum Transmit Window Size (Kbytes) TCP throughput vs M T W S for different 1/a values 60 O Figure 5.14 Chapter 5 Analysis of TCP in a Mobile Environment 88 1/a = 140 sec T C P Reno, 1/b = 5 sec T C P Reno, 1/b = 2 sec T C P Vegas, 1/b = 5 sec T C P Vegas. 1/b = 2 sec o s o 6 0 . Figure 5.15 Maximum Transmit Window Size (Kbytes) TCP throughput vs. MTWS for different 1/b values Figure 5.16 illustrates the throughput comparison of T C P Reno and T C P Vegas, respec-tively, as a function of the M T W S for different P L R s in the bad state within the normal state with 1/a = 140 sec and 1/b = 2 sec at a mobile speed of 65 Km/h. The M T W S has great effects on the throughput of both T C P implementations at different P L R s . At a low P L R in the bad state within the normal state, their throughput increases as the M T W S increases. Beyond a certain M T W S , their throughput gradually decreases because a bigger M T W S creates more packet losses by overloading the network buffers. A t a high P L R , their throughput sharply decreases when the M T W S increases. Because a larger M T W S requires a much longer recovery time from the channel interruptions, a smaller M T W S gives better T C P throughput. Table 5.1 summarizes the optimal M T W S s for both T C P implementations under different conditions and their corresponding throughput. From table 5.1, it seems that a bigger M T W S is more suitable for T C P Reno and T C P Vegas at a relatively low mobility rate, but a smaller M T W S is more suitable for them at a relatively high mobility rate. For a longer 1/b and a high P L R , a small M T W S can improve both T C P throughput performances because of the relatively low P L R per each full sending window size. With a better congestion avoidance mechanism and a faster Chapter 5 Analysis of TCP in a Mobile Environment 89 retransmission mechanism, T C P Vegas, overall, has much better throughput than T C P Reno under most conditions. " 2 3 0 0 pa J3 OO 3 O l O 2 0 3 0 4 0 E Maximum Transmit Window Size (Kbytes) Figure 5.16 TCP throughput vs. MTWS for 1/a = 140 sec Table 5.1 The optimal transmit window sizes for both TCPs under different conditions T C P Reno T C P Vegas Tx window size at mobile speed = 45 Km/h Throughput (Bytes/sec) 4 Kbytes 2191 8 Kbytes 2182 Tx window size at mobile speed = 75 Km/h Throughput (Bytes/sec) 4 Kbytes 2130 4 Kbytes 2168 Tx window size at mobile speed =110 Km/h Throughput (Bytes/sec) 2 Kbytes 2116 4 Kbytes 2144 Tx window size at 1/a = 70 sec Throughput (Bytes/sec) 2 Kbytes 1973 2 Kbytes 1994 Tx window size at 1/a = 140 sec Throughput (Bytes/sec) 4 Kbytes 2130 4 Kbytes 2168 Tx window size at 1/a = 210 sec Throughput (Bytes/sec) 8 Kbytes 2148 32 Kbytes 2181 Chapter 5 Analysis of TCP in a Mobile Environment 90 5.6.2 Effects of Maximum ACK Delay This subsection mainly studies the effects of the M A D on T C P throughput performance in -B R C I I environment. A l l the other parameters are fixed by varying the M A D for different mobile speeds and different M T W S s . Between handoffs, an average P L R of 0.05 in the bad state within the normal state is assumed. The M S S is assumed to be 512 bytes. The radius of the radio cells and the area of the handoff region are fixed at 2 K m and 100 m, respectively. In the first scenario, the M T W S is assumed to be 16 Kbytes, but the mobile speed is varied. In the second scenario, the mobile speed is fixed at 75 Km/h , but the M T W S is varied. Figure 5.17 shows the throughput of T C P Reno and T C P Vegas as a function of the M A D at different mobile speeds. Figure 5.18 and 5.19 show that the throughput of T C P Reno and T C P Vegas as a function of the M A D for differ-ent M T W S s . is> 2 1 O O 3 (B 2 0 S O JS 00 3 O T C P Vegas, Mobile speed = 45 Km/h T C P Vegas, Mobile speed = 75 Km/h T C P Vegas, Mobile speed = 110 Km/h T C P Reno. Mobile speed = 45 Km/h T C P Reno, Mobile speed = 75 Km/h T C P Reno, Mobile speed = 110 Km/h 3 4 5 6 Maximum A C K Delay (msec) Figure 5.17 TCP throughput vs M A D at different mobile speeds Figure 5.17 shows that the M A D greatly affects the throughput of both T C P Reno and T C P Vegas at different mobile speeds. Their throughput increases as the M A D increases; beyond a certain M A D , their throughput gradually decreases. At a relatively high mobile speed, it seems Chapter 5 Analysis of TCP in a Mobile Environment 91 that a longer M A D gives better throughput for both T C P implementations; On the other hand, a shorter M A D gives better throughput of that at a relatively slow mobile speed. It is because a longer M A D reduces the data transmission rate at the sender, particular in a high mobil i ty environment, in order to reduce P L R during handoffs. At a lower mobile speed (e.g. 45 Km/h), a shorter M A D not only can increase the data transmission rate at the sender but also have fast error recovery response, thus improving the throughput of both T C P Reno and T C P Vegas. Chapter 5 Analysis of TCP in a Mobile Environment 92 Figure 5.18 and figure 5.19 show that a smaller window size gives better throughput for both T C P Reno and T C P Vegas over a whole range of the M A D s . When the M A D is increased, their throughput is increased. Beyond a certain A C K delay, their throughput is gradually decreased because the congestion window in the sender grows too slowly to bring the throughput back to the norminal level before the next handoff occurs. A smaller M T W S and a shorter M A D , give a much higher throughput for both T C P implementations. However, with a larger M T W S , particularly for T C P Reno, a longer M A D is needed for good throughput performance because the longer M A D reduces the data transmission rate at the sender. The optimal M A D s and the corresponding T C P throughput under different conditions are summarized in Table 5.2. From table 5.2, T C P Vegas, overall, has better throughput performance than T C P Reno. Table 5.2 The optimal M A D s for TCP Reno and T C P Vegas TCP Vega T C P Reno A C K delay at mobile speed = 45 Km/h and transmit window size =16 Kbytes 1 msec 2 msec Throughput (Bytes/sec) 2165 2039 A C K delay at mobile speed = 75 Km/h, 110 Km/h and transmit window size = 16 Kbytes 3 msec 3 msec Throughput (Bytes/sec) 2101, 2094 2020, 2009 A C K delay, transmit window size = 4 bytes 1 msec - 2 msec Throughput (Bytes/sec) 2136 2019 A C K delay, transmit window size = 16 bytes 2 msec 3 msec Throughput (Btytes/sec) 2080 2019 5.6.3 Effects of Maximum Segment Size Since the P L R in the bad state within the normal state is directly depended on the B E R and segment size, the net effect of the segment size for the specific B E R is to change the P L R in the bad state, which effects on T C P have been considered in figure 5.5 in section 5.4. If the segment Chapter 5 Analysis of TCP in a Mobile Environment 93 size increases at a specified B E R , the P L R is increased almost linearly. However, in this wireless environment, propagation conditions may render the wireless channel unavailable for a short period of time or mobile users needs to be handoffed when they move out of one cell to another cel l . Therefore, the P L R over a mobile channel not only depends on the P L R in the bad state within the normal state but also depends on the mobility rate and the handoff delay. In order to maximize T C P throughput over the mobile radio channel, this subsection investigates the effects of the M S S on the throughput of T C P Reno and T C P Vegas in the M R C I I environment at different mobile speeds for different P L R s in the bad state. A l l other simulation parameters are fixed, but the mobile speed is varied for different M S S s in the first scenario and the 1/a is varied for different M S S s in the second scenario. The M T W S is assumed to be 8 Kbytes. The M A D is assumed to be 1 msec. The B E R in the bad state within the normal state is fixed at l x l O " 6 (a typical B E R in wireless channels). In the first scenario, the radius of the radio cells and the area of the handoff region are fixed at 2 K m and 100 m, respectively. In the second, the mobile speeds is fixed at 75 Km/h . Figure 5.20 shows the throughput of T C P Reno an T C P Vegas as a function of the M S S at different mobile speeds. Figures 5.21 and 5.22 show the throughput of T C P Reno and T C P Vegas, respectively, as a function of 1/a for different M S S s . Figure 5.20 shows that the M S S has a greater effect on the throughput of T C P Vegas than on that of T C P Reno at different mobile speeds. At the mobile speeds of 45 Km/h and 75 Km/h , the throughput of T C P Reno is gradually increased when M S S increased to a certain value and then decreases. A t a mobile speed of 110 Km/h , the throughput of T C P Reno is hardly changed when the M S S is increased. A relatively small M S S such as 256 bytes, gives the worst throughput for T C P Vegas at different mobile speeds. At a lower mobile speed, the throughput of T C P Vegas is first sharply increased and then gradually decreased when the M S S is increased. A t a mobile speed of 110 Km/h , a larger M S S seems to give a higher throughput in T C P Vegas. Chapter 5 Analysis of TCP in a Mobile Environment 1 9 0 0 1 8 0 0 £• ^ 1 T O O • 1 GOO • T C P Reno, Mobile speed = 45 Km/h T C P Reno, Mobile speed = 75 Km/h T C P Reno. Mobile speed =110 Km/h T C P Vegas, Mobile speed = 45 Km/h T C P Vegas, Mobile speed = 75 Km/h T C P Vegas, Mobile speed = 110 Km/h S 1 S O O \ -S O O 1 o o o M a x i m u m s e g m e n t s i z e ( B y t e s ) Figure 5.20 TCP throughput vs the MSS for different mobile speeds S O 1 0 0 1 2 0 1 4 0 A v e r a g e t i m e b e t w e e n h a n d o f f s ( s e c ) Figure 5.21 The throughput of TCP Reno vs 1/a for different MSSs MSS = 256 bytes M S S = 512 bytes - — M S S = 1024 bytes MSS = 1460 bytes S O 1 0 O 1 2 0 1 4 0 A v e r a g e t i m e b e t w e e n h a n d o f f s ( s e c ) Figure 5.22 The throughput of TCP Reno vs l/a for different MSSs Chapter 5 Analysis of TCP in a Mobile Environment 93 Figures 5.21 and 5.22 show that the throughput of T C P Reno and T C P Vegas is increased when 1/a is increased for different MSSs . However, too small a M S S may significantly reduce the amount of data sent per each full sending window, thus degrading T C P throughput. On the other hand, too large a M S S may increase the probability of error per packet, thus also degrading T C P throughput. A t the mobile speed = 75 Km/h and 1/b = 3.5 sec, both T C P implementations with a 512 bytes M S S have better throughput than that with the other M S S s ; whereas with a 256 byte M S S , particular for T C P Vegas, gives the worst throughput over the whole range of the 1/a values. Table 5.3 compares the throughput of both T C P Reno and T C P Vegas as a function of the M S S for different B E R s in the bad state, at 1/b = 2 sec, 1/a = 60 sec, and the mobile speed = 75 Km/h . Table 5.3 The Throughput of TCP Reno and TCP Vegas (Bytes/sec) Maximum segment size TCP Reno BER in the bad state = l x l O 6 TCP Vegas BER in the bad state = lxlO"6 TCP Reno BER in the bad state = 5xl0"s TCP Vegas BER in the bad state = 5xl0"s 256 1771.4 1361.8 1570.4 843.1 384 1689.8 1760.2 1345.3 1333.7 512 1887.5 2003.7 1270.7 1315.2 640 1851.1 2083.2 1119.5 1178.8 768 1840.6 2094.8 1178.2 1249.5 896 1912.9 2168.1 955.16 1090.9 1024 2021.1 2188.5 901.9 925.3 1460 2020.1 2189.4 763.3 795.4 Table 5.3 shows that a larger M S S has a higher throughput performance in both T C P Reno and T C P Vegas for a relatively low P L R in the bad state. It is because a larger M S S can carry more data for every round-trip time before the next channel interruption occurs. However, at a higher P L R , a smaller M S S seems to give higher throughput because the P L R affects not only data packets but also A C K packets. Because of the detail calculation of RTT and RTO for every T C P segment in T C P Vegas, T C P Reno has much better throughput than T C P Vegas when a M S S ChapterS Analysis of TCP in a Mobile Environment 96 of 256 Kbytes or less is used. Otherwise, T C P Vegas has better throughput than T C P Reno for the whole range of the MSSs at different mobile speeds for different 1/a values. 5.7 Optimization and Discussion In order to improve throughput performance of T C P Reno and T C P Vegas in the M R C I I environment under most conditions, this section provides a detailed evaluation for jointly optimiz-ing several T C P parameters including the M T W S , the M A D , and the M S S . From the simulation results in section 5.6, an upper bound and a lower bound of the T C P parameters are obtained and suggested for both T C P implementations under different conditions. In this section, all other simulat ion parameters are fixed, but the above T C P parameters are varied under different conditions. The R T T coefficient deviation is assumed to be 4. The thresholds ( a and (3) in T C P Vegas are assumed to be 1 and 3. The B E R in the bad state within the normal state is fixed at l x l O " 6 . The area of the handoff region is fixed at 100 m. When the mobile host moves from one cell to another cell, the wireless channel is assumed to be unavailable for 1/b sec every 1/a sec. Figures 5.23-5.26 show the throughput of T C P Reno with different M T W S s and M A D s as a function of 1/a for different M S S s , respectively, at a mobile speed of 75 K m / h . From those figures, for a M S S > 256 bytes, T C P Reno with a 4 Kilobyte M T W S and a smaller M A D seems to have much higher throughput than the other M T W S s over the whole range of the 1/a values. For a M S S of 256 bytes, a 2 Kilobyte M T W S seems to give higher throughput for T C P Reno when 1/a < 180 sec. T C P Reno with a 4 Kilobyte M T W S and a longer M A D seems to have much higher throughput than the other M T W S s at a longer 1/a. However, a 2 Kilobyte M T W S seems to give higher throughput than the other M T W S s at a shorter 1/a. For different M A D s and different MSSs , a 32 Kilobyte M T W S seems to give much less throughput than the other M T W S s over the whole range of the 1/a values. Chapter 5 Analysis of TCP in a Mobile Environment 97 2200 r 2000h 18O0r - 1600h 1400 k 1200h TCP Reno Max. segment size = 256 bytes Mobile speed = 75 Km/h 1000 L M A D = 1 msec M T W S = 2 Kbytes M T W S = 4 Kbytes M T W S = 8 Kbytes M T W S = 16 Kbytes M T W S = 32 Kbytes 40 60 0 100 120 140 160 Average time between handoffs (sec) 180 200 220 2200 2100 2000 1900 2 1800. 1,1700 1600 1500 1400 1300 TCP Reno Max. segment size = 256 bytes Mobile speed = 75 Km/h M A D = 3 msec - X — M T W S = 2 Kbytes M T W S = 4 Kbytes M T W S = 8 Kbytes M T W S = 16 Kbytes - O — M T W S = 32 Kbytes 0 100 120 140 160 Average tine between handoffs (sec) 180 200 220 Figure 5.23 Throughput of TCP Reno with a 256 Kbyte MSS vs. 1/a for different M A D s 2200 r 21001-2000h S 1900H 1 1 1-TCP Reno Max. segment size=512 Mobile speed=75 Km/h 1600h 1500r 1400 L M A D = 1 msec M T W S = 2 Kbytes M T W S = 4 Kbytes M T W S = 8 Kbytes M T W S = 16 Kbytes M T W S = 32 Kbytes 00 120 140 160 lime between handoffs (sec) 180 200 220 22001 2100 2000 1900! -1800 : 1700 1600 1500 1400' n 1 r TCP Reno Max. segment size=512 bytes Mobile speed=75 Km/h M A D = 3 msec M T W S = 2 Kbytes M T W S = 4 Kbytes M T W S = 8 Kbytes M T W S = 16 Kbytes — O — M T W S = 32 Kbytes 80 100 120 140 160 Average tine between handoffs (sec) 180 200 220 Figure 5.24 Throughput of TCP Reno with a 512 byte MSS vs. 1/a for different MADs Chapter 5 Analysis of TCP in a Mobile Environment 98 2200r 2100r 2000F 1900r •o iaoh [2 c 1600t-h 1500r 1400h 1300 r 1200h TCP Reno Max. segment size = 1024 bytes Mobile speed = 75 Km/h 1100L M A D = 1 msec M T W S = 2 Kbytes M T W S = 4 Kbytes M T W S = 8 Kbytes M T W S = 16 Kbytes M T W S = 32 Kbytes 2200 2100 2000 1900 7,1700 1600 1500 1400! 1300 TCP Reno Max. segment size-1024 bytes Mobile speed = 75 Km/h M A D = 3 msec M T W S = 2 Kbytes M T W S = 4 Kbytes M T W S = 8 Kbytes M T W S = 16 Kbytes M T W S = 32 Kbytes 40 60 80 100 120 140 160 180 200 220 20 40 Average lime between handoffs (sec) 0 100 120 140 160 1 80 200 220 Average time between handoffs (sec) Figure 5.26 Throughput of TCP Reno with a 1024 byte MSS vs. 1/a for different M A D s Chapter 5 Analysis of TCP in a Mobile Environment 99 From figures 5.23-5.26, for different M A D s and different M T W S s , a M S S within a range between 512 bytes to 1024 bytes (e.g. 896 bytes in this case) seems to give higher throughput for T C P Reno than other M S S s over the whole range of the 1/a values; whereas a 256 byte M S S seems to give the worst throughput for T C P Reno than other M S S s . For different M S S s and different M T W S s , a 1 mill isecond M A D seems to give higher throughput for T C P Reno than other M A D s particular in a high mobility environment. Over a whole range of 1/a, a 4 Kilobyte M T W S gives reasonably high throughput for T C P Reno for different M S S s and different M A D s . Figure 5.27-5.30 show at a mobile speed of 75 Km/h , the throughput of T C P Vegas with different M T W S s and different M A D s as a function of 1/a for different MSSs , respectively. From those figures, For a shorter M A D and a M S S < 512 bytes, a 2 Kilobyte M T W S gives much higher throughput for T C P Vegas than the other M T W S s over the whole range of the 1/a. However, a 4 Kilobyte M T W S seems to give higher throughput for T C P Vegas for a shorter M A D and a M S S > 512 bytes. For a longer M A D and a longer 1/a, a 4 Kilobyte M T W S gives much higher through-put for T C P Vegas than the other M T W S s for different MSSs . However, a 2 Kilobyte M T W S and a smaller M S S give slightly better throughput for T C P Vegas than the other M T W S s for a shorter 1/a. For different M A D s and different MSSs , a 32 Kbyte M T W S gives the worst throughput for T C P Vegas than the other M T W S s over the whole range of the 1/a. Overall, for different M A D s and different M T W S s , a 896 byte M S S seems to give reason-ably high throughput performance for T C P Vegas over the whole range of the 1/a; a 256 byte M S S seems to give the worst throughput performance for T C P Vegas. For different M S S s and different M T W S s , a 1 millisecond M A D seems to give higher throughput performance for T C P Vegas than other M A D s over the whole range of the 1/a, particular in the high mobility environment. For different M S S s and different M A D s , a 8 Kbyte M T W S give reasonably high throughput perfor-mance for T C P Reno over the whole range of the 1/a. Chapter 5 Analysis of TCP in a Mobile Environment 100 2200r 20001-16O0r-"§.1200^ 1000h 600 400' TCP Vegas Max. segment see = 256 bytes Mobile speed = 75 Km/h M A D = 1 msec M T W S = 2 Kbytes M T W S = 4 Kbytes M T W S = 8 Kbytes M T W S = 16 Kbytes M T W S = 32 Kbytes 0 100 120 140 160 Average time between handoffs (sec) 2200 2000 1800 1600 •1400 »!200 1000 800 600 400 TCP Vegas Max. segment size = 256 bytes Mobile speed = 75 Km/h M A D = 3 msec M T W S = 2 Kbytes M T W S = 4 Kbytes M T W S = 8 Kbytes M T W S = 16 Kbytes M T W S = 32 Kbytes 0 100 120 140 160 Average lime between handoffs (sec) Figure 5.27 Throughput of TCP Vegas with a 256 byte MSS vs. 1/a for different M A D s 1400L 0 100 120 140 160 Average time between handoffs (sec) 180 200 220 1400L 0 100 120 140 160 Average lime between handoffs (sec) 180 200 220 Figure 5.28 Throughput of TCP Vegas with a 512 byte MSS vs. 1/a for different M A D s Chapter 5 Analysis of TCP in a Mobile Environment 101 1400' 1 1 1 1 1 1 1 ' 1 1 HOO1 1 1 1 1 1 1 1 1 1 20 40 60 80 100 120 140 160 180 200 220 20 40 60 80 100 120 140 160 180 _ 200 220| Average time between handoffs (sec) Average time between handoffs (sec) Figure 5.29 Throughput of TCP Vegas with a 896 byte MSS vs. 1/a for different MADs Chapter 5 Analysis of TCP in a Mobile Environment 102 Based on a series of simulations to jointly optimize the T C P parameters for both T C P Reno and T C P Vegas at different mobile speeds and for different 1/a values. Table 5.4 summarizes the suggested optimal values of the chosen T C P parameters for T C P Vegas and T C P Reno under most conditions for the M R C I I environments. Figure 5.31 shows the throughput of both T C P implementations with the optimal T C P parameters as a function of 1/a at a mobile speed = 75 Km/h , 1/b = 3.5 sec and the P L R in the bad state within the normal state = 0.04. Table 5.4 The jointly optimal T C P parameters for TCP Vegas and T C P Reno Max. Max. A C K Segment Size x Window Size Delay (msec) (Bytes) T C P Vegas 8 Kbytes 1 896<x<1024 T C P Reno 4 Kbytes 1 896<x<1024 Average time between handoffs (sec) Figure 5.31 TCP throughput vs. 1/a Compared with figure 5.10, figure 5.31 shows that the throughput of T C P Reno is signifi-cantly improved after joint ly opt imizing the T C P parameters particular in the low mobil i ty environment. However, the throughput of T C P Vegas is slightly improved after optimizing the Chapter 5 Analysis of TCP in a Mobile Environment 103 T C P parameters over the whole range of the 1/a. After jointly optimizing the T C P parameters, T C P Vegas still has better throughput than T C P Reno under most conditions. With the optimal T C P parameters, T C P Vegas has approximately the same throughput as T C P Reno in a relatively high mobility environment; however, T C P Vegas has a much higher throughput than T C P Reno in a relatively low mobility environment. Overall, T C P Vegas has approximately 0.1% to 8% better throughput than T C P Reno when 1/a varies from 35 sec to 200 sec. C h a p t e r 6 S u m m a r y a n d Conc lus ions The interest in the performance analysis of T C P implementations over wireless channels interconnected with the Internet is motivated by the critical need for the desired quality of service in seamless E T E delivery of the data traffic over a H W W network. Previous studies have demonstrated the poor performance of both T C P R F C and T C P Reno in networks with satellite links or with mobile links. This thesis focuses on the performance evaluation of T C P R F C , T C P Reno, and T C P Vegas in networks with satellite links and mobile links interconnected with remote L A N s over the Internet. It also studies the effects of the T C P parameters on the T C P throughput performance, while attempting to improve the T C P throughput performance by jointly optimizing the T C P parameters for both H W W networks. 6.1 Summary of Findings The throughput performance of the different T C P implementations in networks with satellite links and mobile links interconnected with the Internet has been thoroughly analyzed and compared. In the BSII environment, the effects of the network elements including the P L R in the satellite links, the STPs between the clear sky state and the rain fade state and the average E T E delay on T C P throughput has been studied. In the M R C I I environment, the effects of the network elements including the STPs among the three states (the handoff state, the good state and the bad state within the normal state), the mobile fading, mobility rates, and handoff interruptions on T C P throughput has also been studied. The effects of the chosen T C P parameters including the M T W S , the M A D , and the M S S on T C P throughput in both H W W networks have been investi-gated, while the optimal T C P parameters for the different T C P implementations in both H W W networks have also been obtained. 104 Chapter 6 Summary and Conclusions 105 From the simulation results, the chosen T C P parameters strongly affect the throughput of the different T C P implementations in networks with satellite links and mobile links intercon-nected with the Internet with limited resources. Too big M T W S s and too small M A D s increase network congestion and cause network buffer overflow; whereas too small M T W S s and too large M A D s result in inefficient bandwidth utilization. Too large M S S s increase the probability of packet error; whereas too small MSSs cause inefficient bandwidth utilization. In the presence of the long propagation delays and high P L R s of satellite links, the three different T C P implementations cannot have efficient bandwidth usage as T C P throughput oscillates with the congestion window. The P L R in the satellite link, the E T E delay, and the chosen T C P parameters have much greater effect on the throughput of T C P Vegas than on that of T C P Reno and T C P R F C ; whereas those network elements only have a slight effect on the throughput of T C P R F C . The big window option works wel l and significantly improves the throughput for both T C P Reno and T C P Vegas only at a relatively low P L R but seriously degrades the throughput of T C P R F C . Because of the limitations of the Internet and the trade-off of a bigger M T W S (creating its own losses by overloading the network), both T C P Reno and T C P R F C prefer to use a smaller M T W S and a shorter M A D , particular in high P L R networks, in order to reduce the probability of congesting the network and to have a faster error recovery. However, because of its better congestion avoidance mechanism and use of a finer-grained timer to update the current R T O for every segment, T C P Vegas prefers to use a bigger M T W S and a longer M A D in order to send more data packet per full window size and have more accurate information in the returned acknowledgment. A larger M S S can carry more data for every round-trip time and can fill up a long satellite pipe more efficiently, thus improving the T C P throughput in network with satellite links. Chapter 6 Summary and Conclusions ]06 The optimal T C P parameters for the different T C P implementations are summarized in Table 4.4. After jointly optimizing the chosen T C P parameters, T C P R F C is still much inferior in throughput to T C P Reno and T C P Vegas under most conditions; whereas T C P Vegas has much higher throughput than the others. T C P Vegas, overall, has an approximately 23% to 100% better throughput than T C P Reno and 500% to 700% better throughput than T C P R F C as the P L R in the satellite links varies from 0.001 to 0.06. T C P Vegas has nearly 40% better throughput than T C P Reno and more than 700% better throughput than T C P R F C over the whole range of the E T E delays. In the presence of high B E R s , high mobile speeds, intermittent connectivity, and handoff interruptions in the mobile radio channel, both T C P Reno and T C P Vegas cannot perform well in the M R C I I environment because these network elements completely violate most of T C P ' s operational assumptions. The network elements and the chosen T C P parameters have great effects on their throughput performances. A t relatively high mobile speed, they prefer to use a smaller M T W S and a shorter M A D . At a relatively low mobile speeds, they prefer to use a bigger M T W S and a longer M A D . For different 1/a values and different l/b values, a smaller M T W S and a shorter M A D give a higher throughput for both T C P implementations. A larger M S S seems to have relatively high throughput efficiency for both T C P implementations even under high mobility conditions. The jointly optimal T C P parameters for both T C P Reno and T C P Vegas are summarized in Table 5.4. After jointly optiminzing T C P parameters, the throughput of T C P Vegas is overall better than T C P Reno under most conditions. T C P Vegas has an approximately 0.1% to 8% better throughput than T C P Reno when the average time between handoffs varies from 35 sec to 200 sec at the P L R in the bad state = 0.04, the mobile speed = 75 Km/h , and l/b = 3.5 sec. Chapter 6 Summary and Conclusions 107 6.2 Future Investigation Based on the following main reasons, network congestion is unlikely and it is reasonable for a proposed T C P implementation to assume by default that any loss is due to medium errors. First, most of the losses in the Internet occur one packet at a time [32]. Second, the simulation results show that most of the losses are due to media error i f a better congestion avoidance mechanism is used. The coarse-grain R T O and the exponential backoff algorithm seriously degrade the T C P throughput performance. Thi rd , in today's high speed networks, network bandwidth is carefully managed on private links. The proposed T C P (P-TCP) uses the T C P Vegas variant of the Slow Start algorithm and the congestion avoidance algorithm in T C P Vegas. P - T C P uses the fine-grain R T O algorithm in T C P Vegas to retransmit lost segments, but it does not apply the coarse-grain R T O and the exponential backoff algorithm. Simulation results show that P - T C P has an approximately 6% to 20% better throughput than T C P Vegas as P L R in the satellite links varies from 0.005 to 0.05 (see figure 6.1). It also has an approximately 3% to 30% better throughput than T C P Vegas as the P L R in the bad state varies from 0.005 to 0.1 (see figure figure 6.2). However, the multiple packet losses per each full sending window size, still seriously degrade T C P throughput. Therefore, subject to the continued interest in the performance improvement of T C P for networks with satellite l inks and mobile links interconnected with the Internet, further research and more analysis may be needed. The following areas are suggested for further investigation: • A better T C P acknowledgment method in the receiver to inform the sender to send all the lost packets without overflowing the buffers Chapter 6 Summary and Conclusions 108 • A better retransmission method to enable the sender to discriminate packet losses due to the media from packet losses due to congestion. 1 3 0 i 1 1 1 1 1 1 1 r 3 0 I i i i 1 1 : 1 1 1 1 0 . 0 0 5 0 . 0 1 0 . 0 1 5 0 . 0 2 0 . O 2 5 0 . 0 3 0 . 0 3 5 0 . 0 4 0 . 0 4 5 0 . 0 5 Packet Loss Rate in the satellite channel Figure 6.1 The throughput comparison of TCP Vegas and a proposed T C P vs. the PLR in the satellite links 2 5 0 0 r— 1 1 1 1 1 1 1 1000 \-3 0 i ' -j 1 1 1 ' ' ' ' 0 . 0 1 0 . 0 2 0 . 0 3 0 . 0 4 0 . 0 5 0 . 0 6 0 . 0 7 0 . 0 8 0 . 0 9 0 .1 Packet loss rate when the wireless channel staying in the bad state Figure 6.2 The throughput comparison of TCP Vegas and a proposed T C P vs. the PLR in the bad state within the normal state for networks with mobile links Bibliography [1] M . Schwartz, "Network Management and Control Issues in Multimedia Wireless Networks", IEEE Personal Communications, pp. 8-16, June 1995. [2] M . Talla, A . K . Elhakeem and M . Kadoch, "Throughput and Delay Characteristics of A R Q Transport Protocols over L E O and G E O Satellite Networks", Singapore ICCS 94, pp. 410-414, 1994. [3] M . Aghadavoodi Jolfaei, B . Heinrichs, and M . R. Nazeman, "Improved T C P Error Control for Heterogeneous W A N s " , IEEE National Telesystems Conference 94, pp. 257-260, July 1994. [4] Oliver W. W. Yang, Xiao-Xiong Yao, and K . M . S. Murthy, "Modeling and Performance Analysis of File Transfer in a Satellite Wide Area Network", IEEE Journal on Selected Areas in Communications, Vol . 10, No . 2, pp. 428 -436, Feb. 1992. [5] H . Linder, W. Stering, U. Hofmann, "Performance of X T P X and TCP/IP in a satellite environment", Proceedings of 20th Conference on Local Computer Networks, pp. 246-253, October 1995. [6] P. Manzoni , D . Ghosal, G . Seazzi, " A simulation Study of the Impact of Mobi l i ty on T C P / IP", Network Protocol, 1994 International Conference, pp. 196-202, May 1994. [7] R. Caceres and L . Iftode, "Improving the Performance of Reliable Transport Protocols in Mobi le Computing Environments", IEEE Journal on Selected Areas in Communications, pp. 850-857, June 1995. [8] M . C . Chuah, O. C . Yue and A . DeSimone, "Performance of Two T C P Implementations in Mobi le Computing Environments", IEEE Journal on Selected Areas in Commumications, pp. 339-344. June 1995 [9] Z . Haas and Chih-Lin I, "On Handoffs in Packetized Wireless Systems", Globecom '93: IEEE Global Telecommunications Conference, pp. 1808-1814, June 1993. [10] H . Balakrishnan, S. Seshan, and R. H . Katz, "Improving Reliable Transport and Handoff Performance in Cellular Wireless Network", ACM Mobile Computing and Networking Conference, Berkeley, California, pp. 14-15, November 1995. 109 Bibliography 110 [11] P. Karn and C. Partidge, "Improving Round Trip Time Estimates in Reliable Transport Protocols", ACM Transactions on Computer Systems, pp. 364-373, Sept. 1991. [12] Protocol Engines, Inc., " X T P Protocol Definition", Revision 3.6, XTP Forum, Jan. 1992. [13] A . Bakreand B . R. Badrinath, " I -TCP: Indirect T C P for Mobile Host", Technical Report DCS-TR-314, Rutgers University, Oct. 1994. [14] E . Amir , H . Balakrishnan, S. Seshan, and R. H . Katz, "Efficient T C P over Networks with Wireless Links", Proc. ACM Mobicom, Nov. 1995 [15] A . DeSimone, M . C. Chuah, and O.C. Yue, "Throughput Performance of Transport-Layer Protocols over Wireless L A N s " , Proc. Globecom '93, Dec. 1993. [16] H . K . Randy, "Adaption and Mobil i ty in Wireless Information Systems", Proc. ACM Mobicom, Nov. 1995. [17] V. Jacobson, "Congestion Avoidance and Control", Proceeding of ACM SIGCOMM' 88, Aug . 1988. [18] W. Stevens, " T C P Slow Start, Congestion Avoidance, Fast Retransmit, and Fast Recovery Algorithms", Request for Comments 2001, IETF, January 1997. [19] V. Jacobson and R.T. Braden, " T C P Extensions for Long Delay Paths", Request for Comments 1072, IETF. October, 1988. [20] V. Jacobson, R. T. Braden, and D . A . Borman, " T C P Extensions for High Performance", Request for Comments 1323, IETF, 1988. [21] R. Fox, " T C P B i g Window and Nak Options", Request for Comments 1106, IETF, June, 1989. [22] V. Paxson, "Empirically Derived Analytic Models of Wide-Area T C P Coonnection", IEEE Transactions on Networking, Vol . 2, No. 4, August 1994. [23] R. Caceres, P. B . Danzig, S. Jamin, D . J. Mitzel , "Characteristics of Wide-Area TCP/IP Conversations", ACM SIGCOMM' 91, pp. 103-110, May 1991. Bibliography 111 [24] V. Paxson and S. Floyd, "Wide Area Traffic: The Failure of Poisson Modeling", IEEE/ ACM Transactions on Networking, Vol . 3, No. 3. June 1995 [25] K . M . Kha l i l , Y. S. Sun, "Peformance Considerations for TCP/IP in Wide Area Networks", Local Computer Networks, 1994 19th Conference, pp. 166-175, March 1993 [26] L . S. Brakmo and L . L . Peterson, " T C P Vegas: end to end congestion avodiance on a global Intenet", IEEE Journal on Selected Areas in Communications, Vol 13, No. 8. pp. 1465-1480, October 1995. [27] W. R. Stevens, TCP/IP Illustrated, Volume 1: The Protocols. Addison-Wesley, 3rd ed., 1994. [28] J. B . Postel, "Transmission Control Protocol", Request for Comments 793, IETF, Sept., 1981. [29] R. F. Quick, " A n overview of the cellular digital packet data system", Proceedings of the Fourth International symposium on Personal, Indoors and Mobile. Radio Communications, pp. 338-343, Sept. 1993. [30] R. Jain, The art of computer systems performance analysis: techniques for experimental design, measurement, simulation, and modeling, Wiley, pp. 491, 1991. [31] M . Fi l ip , E . Vtlar, "Optimum utilization of the channel capacity of a satellite link in the presence of amplitude scintillations and rain attenuation", IEEE Transactions on communications, vol.38, No. 11, pp. 1958-1965, November 1990. [32] A . Sanghi, A . K . Agrawala, and O. Gudmundsson, " Experimental Assessment of End-to-End Behavior on Internet", Computer Communication Proceeding, IEEE INFOCOM'93, pp. 98-115, Nov. 1993. [33] Cellullar Digital Packet Data System Specification. C D P D Forum, Inc. Release 1.1. Jan. 1995. [34] S. Jangi and L .F . Merakos, "Performance Analysis of Reservation Random Access Protocols for Wireless Access Network," IEEE Transaction on Comm. Vol 42, pp. 1223-1234, A p r i l , 1994. [35] D . Anick, D . Mitra, and M . M . Sondhi, "Stochastic theory of a data-handling system with multiple sources," Bell System Technical Journal, Vol . 61, pp. 1871-1894, 1982 Bibliography 112 [36] W. C. Y. Lee, Mobile Cellular Telecommunications Systems, M c G r a w - H i l l Inc., 2nd ed., 1995. [37] K . Kha l i l , J. Hand, and M . Mariswamy, "Analysis and Traffic Characterization of a Wide Area Network," Proceedings of ICC'93, Geneva, pp. 23-26, May 1993. [38] G . Arnold and T. W. Kao, "Rain Attenuation in E H F Satellite Communications", IEEE Pacific Rim Conference on Communications, Computers, and Signal Processing. Proceedings, pp. 153-156, May, 1995. [39] Wolfhard J. Vogel, Geoffrey W. Torrence, and Jeremy E . Allnutt, "Rain Fades on Low Elevation Angle Earth-Satellite Paths: Comparative Assessment of the Austin, Texas, 11.2-GHz Experiment", Poceedings of the IEEE, Vol . 81, No. 6, pp. 885-896, June 1993 Appendix A. List of Abbreviations and Acronyms A C K Acknowledgment A R Q Automatic Repeat Request B E R Bit Error Rate B S Base Station BSII Broadband Satellite Interworking with the Internet C D M A Code Division Multiple Access C D P D Cellular Digital Packet Data E T E End-to-End F R F R Fast retransmit and fast recovery F T P File Transfer Protocol G E O Geosychronous H W W Heterogeneous Wireless and Wired IP Internet Protocol I C M P Internet Control Message Protocol I G M P Internet Group Management Protocol L A N Local Area Network L E O Low Earth Orbit N N T P Network News Transfer Protocol M A D Maximum A C K delay M R C I I Mobile Radio Channel Interworking with the Internet M S S Maximum Segment Size M T W S Maximum Transmit Window Size Appendix A. List of Abbreviations and Acronyms 114 M T U Maximum Transfer Unit M D I S Mobile Data Intermediate System N A K Negative Acknowledgment P L R Packet Loss Rate R T T Round Trip Time R T O Retransmission Time Out T C P Transmission Control Protocol T D M A Time Division Multiple Access U D P User Datagram Protocol Appendix B. Opnet Simulation Models Mobile Host Loss State Haocidoff S t» te Figure B . 1. The heterogenous wireless and wired simulation envrionment model Figure B.2 A End Host Model Figure B.3 Router and Gateway model 115 Appendix B. Opnet Simulation Models 116 F igure B .4 The T C P appl icat ion process mode l (! serverjmsy && insert_ok) (! Q U E U E _ E M P T V ) F igure B.5 The IP and state process models 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0065315/manifest

Comment

Related Items