UBC Faculty Research and Publications

Inferring models of concurrent systems from logs of their behavior with CSight Beschastnikh, Ivan; Brun, Yuriy; Ernst, Michael D.; Krishnamurthy, Arvind Feb 28, 2014

You don't seem to have a PDF reader installed, try download the pdf

Item Metadata


52383-tr_paper.pdf [ 637.96kB ]
JSON: 52383-1.0051021.json
JSON-LD: 52383-1.0051021-ld.json
RDF/XML (Pretty): 52383-1.0051021-rdf.xml
RDF/JSON: 52383-1.0051021-rdf.json
Turtle: 52383-1.0051021-turtle.txt
N-Triples: 52383-1.0051021-rdf-ntriples.txt
Original Record: 52383-1.0051021-source.json
Full Text

Full Text

University of British Columbia Technical ReportInferring Models of Concurrent Systemsfrom Logs of their Behavior with CSightIvan Beschastnikh , Yuriy Brun , Michael D. Ernst , Arvind KrishnamurthyDepartment of Computer Science School of Computer Science Computer Science & EngineeringUniversity of British Columbia University of Massachusetts University of WashingtonVancouver, BC, Canada Amherst, MA, USA Seattle, WA, USAbestchai@cs.ubc.ca, brun@cs.umass.edu, {mernst, arvind}@cs.washington.eduABSTRACTConcurrent systems are notoriously difficult to debug and understand.A common way of gaining insight into system behavior is to inspectexecution logs and documentation. Unfortunately, manual inspectionof logs is an arduous process, and documentation is often incompleteand out of sync with the implementation.To provide developers with more insight into concurrent systems,we developed CSight. CSight mines logs of a system?s executions toinfer a concise and accurate model of that system?s behavior, in theform of a communicating finite state machine (CFSM).Engineers can use the inferred CFSM model to understand com-plex behavior, detect anomalies, debug, and increase confidence inthe correctness of their implementations. CSight?s only requirementis that the logged events have vector timestamps. We provide a toolthat automatically adds vector timestamps to system logs. Our toolprototypes are available at http://synoptic.googlecode.com/.This paper presents algorithms for inferring CFSM models fromtraces of concurrent systems, proves them correct, provides an im-plementation, and evaluates the implementation in two ways: byrunning it on logs from three different networked systems and viaa user study that focused on bug finding. Our evaluation finds thatCSight infers accurate models that can help developers find bugs.1. INTRODUCTIONWhen a system behaves in an unexpected manner, or when adeveloper must make changes to legacy code, the developer facesthe challenging task of understanding the system?s behavior. To helpwith this task, developers often enable logging and analyze runtimelogs. Unfortunately, the size and complexity of logs often exceed ahuman?s ability to navigate and make sense of the captured data.One promising approach to help developers is model inference.The goal of a model-inference algorithm is to convert a log of sys-tem executions into a model, typically a finite state machine, thataccurately and concisely represents the system that generated thelog. Numerous model-inference algorithms and tools exist to helpdebug, verify, and validate sequential programs [11, 43, 44, 10].Unfortunately, this rich prior body of work is not directly applica-ble to distributed or concurrent systems. This is because a commonassumption made in model inference is that the underlying set of ex-ecutions is totally ordered ? for every pair of events in an execution,one precedes the other. This assumption is crucial to the inferencealgorithms? correctness, and running them on logs that are not totallyordered results in inaccurate models. Unfortunately, this assumptiondoes not hold for concurrent systems, such as networked systems, forwhich events at different nodes may occur without happens-beforerelationships [40]. Additionally, most model-inference algorithmsinfer finite state machine (FSM) models, which are inappropriate formodeling multi-process implementations.This paper describes a new model-inference technique and acorresponding tool, called CSight (for ?concurrency insight?), whichinfers a communicating finite state machine (CFSM) [14] model ofthe processes that generated the log. CSight can be applied to logs ofdistributed systems, protocol traces, traces of AJAX events in a web-browser, and other concurrent behavior traces. CSight models havemultiple developer-oriented uses ? for example, developers caninspect, query, and check CSight?s models against their own mentalmodel of the system to find bugs. These uses are the focus of ourpresent work. However, we believe that CSight-generated modelshave numerous other applications, such as model-based testing ofconcurrent systems, and automated detection of anomalous behavioras systems are exposed to new workloads and environments.This paper presents the following four contributions:1. A novel model-inference algorithm to infer a CFSM modelfrom a log of vector-timestamped concurrent executions (Sec-tion 3), and a corresponding tool called CSight.2. A proof that the CFSM models generated by CSight are accu-rate, in the sense that they satisfy several key properties of theinput log (Section 4).3. A tool for automatically augmenting existing process-locallogging mechanisms with vector-timestamping logic. Thismakes it easy to apply CSight to existing systems (Section 5).4. A two-pronged evaluation of CSight (Section 6):- We applied CSight to logs of three systems ? the stop-and-wait protocol, the opening/closing handshakes of TCP, andthe replication strategy in the Voldemort [57] distributedhash table. CSight was effective for uncovering the truemodel for each of these systems.- We performed a user study with a class of 39 undergraduatesto evaluate the efficacy of CFSM models in finding bugs.We found that CFSM models are just as useful in findingimplementation bugs as time-space diagrams, a popularalternative for visualizing concurrent executions.We start with an overview of the CSight algorithm (Section 2).2. CSIGHT OVERVIEWCSight?s input is a log of observed system execution events, and itsoutput is a model that describes the concurrent system that generatedthe log. An input log consists of execution traces of the system. Atrace is a set of events, each of which has a vector timestamp [25,46]. Vector time is a standard logical clock mechanism that provides1University of British Columbia Technical Reportsend(x)M!x,oA?ack,osend(x)M!x,eA?ack,esend(x)M!x,oA?ack,osend(x)M!x,eA?ack,e1,02,03,34,35,36,67,68,69,910,911,912,12MAs0s1send(x)s2M!x,otimeoutA?ack,es3A?ack,os4send(x)s5M!x,eA?ack,eA?ack,otimeout(b.1) Output model (Sender)A?ack,er0r1M?x,or2recv(x)M?x,or3r4M?x,er5recv(x)A!ack,eM?x,eM?x,eM?x,oA!ack,o(b.2) Output model (Receiver)(a) Input logSender Receiver2,12,22,35,45,55,68,78,88,911,1011,1111,12M?x,orecv(x)A!ack,oM?x,erecv(x)A!ack,eM?x,orecv(x)A!ack,oM?x,erecv(x)A!ack,eFigure 1: Example input and output of CSight for the stop-and-wait protocol (SAW) [55]. (a) Example input log with a single trace oftwo processes running SAW. The two integers at the beginning of each event in the log are the event?s vector clock timestamp. (b) TheCSight-derived CFSM model of SAW, derived from a log with additional traces (not-shown), consisting of (b.1) the sender processmodel and (b.2) the receiver process model.In SAW, the sender transmits messages to the receiver using channel M, and the receiver replies with acknowledgments throughchannel A. Notation Q!x means enqueue message x at tail of channel Q, and event Q?x means dequeue message x from the head ofchannel Q. The event send(x) is a down-call to send x at the sender, and recv(x) is an up-call at the receiver indicating that x wasreceived. In the SAW implementation, x is a variable that stands for an arbitrary message string. CSight, however, treats the loggedevent instances, like ?send(x)? and ?timeout?, as strings and does not interpret them. The timeout event at the sender triggers amessage re-transmission after some internal timeout threshold is reached. The ?alternating bit? is associated with each message andis appended to a message before it is sent. For example, the first (and every odd) message sent by the sender is represented as x,o (ofor odd) while every even message sent by the sender is encoded as x,e (e for even).a partial order of events in the system. Section 5 describes a toolthat automatically adds vector time tracking to existing systems.Figure 1(a) shows an example input log generated by two pro-cesses executing the stop-and-wait protocol [55]. In this protocol, asender process communicates a sequence of messages to a receiverprocess over an unreliable channel. The receiver must acknowledgean outstanding message before the sender moves on to the nextmessage. If a message is delayed or lost, the sender retransmits themessage after a timeout.Figure 1(b) shows CSight?s output ? a communicating finitestate machine [14] (CFSM) model. A CFSM models multiple pro-cesses, or threads of execution, each of which is described by a finitestate machine (FSM). In the standard CFSM formalism, processescommunicate with one another via message passing over reliableFIFO channels. However, unreliable channels can be simulated byreplacing each unreliable channel with a lossy ?middlebox? FSMthat non-deterministically chooses between forwarding and losing amessage. The model in Figure 1(b) handles message loss, but thelossy middlebox is not shown in the diagram.We use the CFSM formalism because it is similar to the widelyknown FSM formalism. CFSMs are well-established in the for-mal methods community, and we believe (and empirically verifyin Section 6) that a CFSM is intuitive and sufficiently simple fordevelopers to comprehend. For example, a single-process FSMin a CFSM can be inspected and understood without needing tounderstand the activity of other processes in the system.CSight infers a CFSM (e.g., Figure 1(b)) by transforming theinput log (e.g., Figure 1(a)) through a series of representations andanalyses. Figure 2 details this process. The CSight process has threekey stages:Stage 1: temporal property mining. CSight mines temporalproperties, or invariants, from the log (steps 1 and 2 in Figure 2).For example, for the log in Figure 1(a), CSight would mine theproperty ?event instance string send(x) always precedes the eventinstance string receive(x)?. Section 3.2 describes all the types ofproperties CSight mines. CSight uses these properties to judge theaccuracy of the current model, as it refines the model to accuratelydescribe the log.Stage 2: create initial model. CSight uses the log to build acompact model (steps 3 and 4 in Figure 2). This model gen-eralizes all the executions in the log, and it is inaccurate in that itadmits executions that violate the properties mined from step one.To build this small model CSight assumes that whenever an eventof a particular type executes, the system must be in a unique stateassociated with events of that type. For example, every time the loghas a send(x) event, CSight assumes the system was in the exactsame send(x) state. This creates an overly permissive but highlycompact model.Stage 3: refine the model. CSight gradually changes the initialcompact model from stage 2 into a larger model that satisfies all ofthe mined properties from stage 1 (steps 5 and 6 in Figure 2). Toaccomplish this, CSight checks which of the mined properties arenot satisfied by the present model. A property is not satisfied if themodel allows an execution that violates the property. CSight checksproperties with model checking, which either guarantees that theproperty is satisfied by all modeled executions, or finds a counter-example execution that violates the property. If a counter-exampleexists, CSight refines the model to eliminate that counter-exampleusing the CEGAR approach [16]. CSight repeats the model-checkingand refinement loop until all of the mined properties are satisfied, atwhich point it outputs the resulting model.The next section formally describes the CSight process and detailseach of the steps in Figure 2.3. FORMAL DESCRIPTION OF CSIGHTIn Section 4, we will prove three important properties aboutCSight?s model-inference process:- Inferred model fits the input log. The final model accepts all theobserved traces in the input log (see Theorem 1).- Refinement always makes progress. Every iteration of the re-finement process (stage 3 in Section 2) makes progress towardsatisfying all of the mined invariants (see Theorem 2).- Inferred model satisfies mined invariants. Every trace acceptedby the final model behaves like the observed traces; that is, everyinvariant that is true of the observed traces is also true of everyaccepted trace (see Theorem 3).2University of British Columbia Technical ReportLogRegularExpressionsTemporal InvariantsSystemTracesConcrete FSMAbstractFSMCFSMInvariantCounter-exampleModel checkingLog parsingTrace linearizationInvariantminingState abstractionProcess projectionUser inputsCSight outputAllinvariantssatisfied?YesNoModelrefinementChannelDefinitionsFinalCFSM Model134B562Validated TemporalInvariantsAInvariantvalidationFigure 2: CSight process flow chart. Section 3 describes thenumbered steps. Steps A and B are distinguished with lettersas they do not appear in [8].To enable proving these properties, we must first formalize theproblem domain. A reader who wishes to only get an intuitiveunderstanding of the approach can skim this section.We start by defining CSight?s inputs (Section 3.1) and the invari-ants that CSight mines (Section 3.2). Then, we specify how CSightconverts a log into a concrete FSM (Section 3.3) and how CSightabstracts the concrete FSM into the initial abstract FSM (AFSM)(Section 3.4). Finally, we describe how CSight model-checks andrefines the AFSM to satisfy the valid invariants, and how the AFSMis converted into a CFSM model (Section 3.5).First, we describe the notation used in the rest of the paper. Givenan n-tuple t, let |t| = n, and for all 1 ? i ? n, let t[i] refer to theith component of t; for i > n, let t[i] = ?. We write t ?? = t ? t ? todenote concatenation of two tuples: t ?? has length |t|+ |t ?|, and forall 1? i? |t|, t ??[i] = t[i] and for all 1? i? |t ?|], t ??[|t|+ i] = t ?[i]. Wecall t a prefix of t ?? iff ?t ? such that t ?? = t ? t ?. The projection functionpi maps a tuple t and a set S to a tuple t ? such that t ? is generated fromt by deleting all components of t that are not in S. Finally, we usethe symbol ? to denote variables that represent abstract objects in themodeling process and to differentiate these from concrete featuresof the input log. For example, for a logged (concrete) event instancee the corresponding abstract event type is denoted as e?.3.1 Expected log inputTo use CSight, the user must provide an input log, a set of user-defined regular expressions, and a set of channel definitions (step1 in Figure 2). The regular expressions determine which log linesare parsed (and which are ignored), what part of a line correspondsto a vector timestamp, and which local, send, or receive event theline represents. The channel definitions are used to associate sendand receive events with inter-process channels.Currently, CSight assumes that the input log contains only com-plete, non-erroneous executions. We plan to lift this assumption inour future work.The rest of this section formally describes the structure of theinput log. We assume that an input log L is produced by a systemcomposed of h processes, indexed from 1 to h. The log containsmultiple system traces, each of which represents a single concurrentexecution of the system. A system trace consists of a set of eventinstances logged by different processes and a happens-before rela-tion [40] (a strict partial ordering over event instances). For example,the log in Figure 1(a) contains a single system trace; in this trace thevector timestamps encode the partial order.Definition 1 (System trace). A system trace is the pair S = ?T,??,where T = ?Ti is the union of a set of process traces (see Def. 2),one per process; and ? is the happens-before relation. This relationpartially orders event instances at different processes and totallyorders event instances (see Def. 3) in each process trace Ti accordingto their positions in the trace. That is, ?ei = ?e?, i,k? ? Ti, e?i =?e??, i,k?? ? Ti, ei ? e?i ?? k < k?.S must satisfy three communication consistency constraints: (1)every sent message is received, (2) only sent messages are received,and (3) sent messages on the same channel must be received in FIFOorder. The validity of the communication consistency constraints #1and #3 depends on the underlying message passing protocol.1We express the above three constraints with a bijection betweenmessage send and message receive event instances for every pair ofprocesses. For all i, j, let Si j ? T be the send events along channelci j, and let Ri j ? T be the receive events along channel ci j. Then?i, j,? a bijection f : Si j? Ri j, such that:1. f (s) = r =? s? r2. ?s1,s2 ? Si j,s1 ? s2 =? f (s1)? f (s2).A system trace is composed of multiple process traces, and eachprocess trace is the set of event instances generated by a specificprocess. We assume that each event instance is given a uniqueposition k in a consecutive order, from 1 to the length of the trace.Definition 2 (Process trace). For the process i, a process trace is aset Ti of event instances, such that ?k ? [1, |Ti|], ??e?, i,k? ? Ti, and?e?1, i,k1? ? Ti and ?e?2, i,k2? ? Ti =? k1 = k2.The execution of each process generates a sequence of eventinstances, each of which has an event type from a finite alphabet ofprocess event types. These event types can be local events, messagesend events, or message receive events.Definition 3 (Event instance). For a process i, an event instanceis a triple e = ?e?, i,k?, where e? is the event type of e, and k ? 1 isan integer that uniquely indicates the order (position) of the eventinstance, among all event instances generated by process i. Whenthe value of k is not important, we denote this triplet as ei.Finally, we assume a fixed set of channels, each of which is usedto connect one sender process to one receiver process. Each of thesend and receive event types is associated with a channel.Definition 4 (Channel). A channel ci j is identified by a pair ofprocess indices (i, j), where i 6= j and i, j ? [1,h]. Indices i and jdenote the channel?s sender and receiver process, respectively.We use the standard notation ! to denote send events, ? to denotereceive events, and use labels for channels. For example, in Figure 1,the event M!x,e is a send of message x,e on channel M.1In this paper we focus on TCP, which satisfies both assumptions.Generally, TCP is used by the complex systems that CSight targets.Removing these assumptions is future work (see Section 8).3University of British Columbia Technical Report3.2 Invariant miningCSight uses the log to mine a set of temporal invariants ? lineartemporal logic expressions ? that relate events in the log (step 2 inFigure 2). These invariants (Def. 5) are true for all of the observedexecution traces. CSight guarantees that the final inferred CFSMmodel satisfies all the mined invariants.Definition 5 (Event invariant). Let L be a log, and let a?i and b? jbe two event types whose corresponding event instances, ai andb j, appear at least once in some system trace in L. Then, an eventinvariant is a property that relates a?i and b? j in one of the followingthree ways.a?i? b? j : An event instance of type a? at host i is always followedby an event instance of type b? at host j. Formally:??T,?? ? L,?ai ? T,?b j ? T,ai ? b j.a?i 6? b? j : An event instance of type a? at host i is never followed byan event instance of type b? at host j. Formally:??T,?? ? L,?ai ? T, 6 ?b j ? T,ai ? b j.a?i? b? j : An event instance of type a? at host i always precedes anevent instance of type b? at host j. Formally:??T,?? ? L,?b j ? T,?ai ? T,ai ? b j.For example, one invariant of the stop-and-wait protocol modelin Figure 1 is M?m-0 ? A?a-0. The invariant types and the cor-responding mining algorithms are described in more detail in [9].The above invariant types are also exactly the most frequently ob-served specification patterns formulated by Dwyer et al. [24] forserial systems.2 In our experience, these invariants were sufficientfor capturing key temporal properties of the systems that producedthe logs we considered.3.3 Deriving a concrete FSMIn step 3 of Figure 2, CSight creates a concrete FSM that acceptsall possible linearizations of the partially ordered system traces in thelog. The concrete FSM model closely fits the logged observations,but it is not a concise model. Later, we will show how CSight usesabstraction to make the concrete FSM concise.3.3.1 Concrete stateTo define a concrete FSM we need to introduce notions of statethat describe the concrete observations in the log. For this, we willdefine process states, channel states, and system states.Definition 6 (Local process state). Each process begins executionin an initial state, q?i , and after executing a sequence s of process ievent instances, the process enters state qsi . More formally, let Li bethe set of process i traces in a log L, then the set of process i localstates is Qi:Qi = {q?i }?{qsi | ?t ? Li,s is a prefix of t}We call qsi a terminal state for process i if and only if s ? Li.Now, we define the global process state and global channel statethat together make up the system state.Definition 7 (Global process state). A global process state q =?q1, . . . ,qh? is a h-tuple that represents the state of all processes inthe system. That is, q ? Q = Q1??? ??Qh, with qi ? Qi denoting astate at process i.2Scope is constrained to a trace (i.e., global scope). The translationis not one-to-one: a?? b? is Dwyer et al.?s Existence pattern when a?is the start event that precedes every trace, and is otherwise Dwyeret al.?s Response pattern. Another example is ?b?, a? ? b?, which isDwyer et al.?s Universality pattern.A channel contains all sent messages not yet received. For eachchannel ci j, a (possibly empty) finite set of messages Mi j are theonly messages that can be sent and received on ci j.Definition 8 (Channel state). The channel state wi j of a channel ci jis a tuple of variable length whose entries are messages that can besent and received along ci j. That is, wi j ? (Mi j)?.Definition 9 (Global channel state). A global channel state w is aset of channel states for all channels in the system. More formally,w = {wi j | wi j ? (Mi j)?}. We reuse ? to also stand for a globalchannel state with all channels empty. Further, we denote the set ofall possible global channel states M.Finally, we represent the system state as a pair of global processstate and global channel state, ?q,w? ? Q?M.3.3.2 TransitionsNext, we specify how a sequence of event instances impacts thestate of the system. For this, we define a process transition function,?i, which maps a process i state and an event instance to a new state.Definition 10 (Process transition function). Let Ei be the set of allprocess i event instances in a log L and let Li be the set of process itraces in L. Then, the process transition function for a process i is?i : Qi?Ei? Qi, such that? ?i(qsi ,ei) = qs?eii ?? ?t ? Li,(s ? ei) is a prefix of t.As an example that illustrates ?i, assume that states are the nat-ural numbers: Qi = N, and event instances Ei are totally ordered:?g : Ei?N. Then, we can define ?i(qi,ei) as the number that isformed by concatenating qi and g(ei).Notice that ?i has a distinguishing property ? two local processstates are different if they were generated by two distinct sequencesof events. Or, more formally:1. ei 6= fi ?? ?i(qi,ei) 6= ?i(qi, fi)2. qi = q?i ?? ?i(qi,ei) = ?i(q?i,ei)3.3.3 The concrete FSMCSight uses the system traces parsed from a log L to construct aconcrete FSM FL (step 3 in Figure 2). This FSM represents theobserved, concrete executions of the whole system. Each state in theconcrete FSM is a tuple of the individual process states and channelcontents (for all processes and channels in the system). A channel?scontent is computed from the sequence of observed message sendsand receives in the trace. Process states, however, must be inferred,and are assumed to be uniquely determined by the process history.In other words, for a specific sequence of events at a process, CSightcreates a single unique anonymous process state. Figure 3 illustrateshow the anonymous process states and the corresponding concreteFSM are derived from a set of input system traces.A key property of FL is that it accepts all linearizations of alltraces in L, as well as all possible traces that are stitchings of differentconcrete traces that share identical concrete states.Definition 11 (Concrete FSM). Given a log L, a concrete FSM FLfor L is a tuple ?S,sI ,E,?,ST ?.? The states of FL are system states: S = Q?M? There is a unique initial system state: sI = ?[q?1, . . . ,q?h],??? E = {e | ??T,?? ? L,e ? T}? The transition function ? is the composition of the individ-ual process transition functions, except that ? also handlescommunication events.? : Q?M?E? Q?M?(?q,w?,ei) = ?q?,w??, where:4University of British Columbia Technical Report q?i = ?i(qi,ei), and q?j = q j if j 6= i. w?i j =?????????wi j ?m ei = ci j!mtail ei = ci j?m, wi j = m ? tailundefined ei = ci j?m, wi j[1] 6= mwi j otherwise? The terminal states have empty channels, and each processis in a terminal state that was derived by executing all of theevent instances for that process in some system trace in L.ST = {?q,?? | ?i,qi is terminal}3.3.4 Validating the mined invariantsAlthough the invariants CSight mines (Section 3.2) are true of theinput traces, in some cases it is not always possible to construct aconcrete FSM model that satisfies the invariants and accepts all ofthe input traces. This situation is caused by the incompleteness ofthe log.Consider Figure 3, which details the construction of a concreteFSM from three input system traces. The traces were generatedby a system with two processes, p1 and p2, with a single channelfrom p1 to p2. The mined invariant b1 6? x2 is true of the log,but it is not valid in the concrete FSM. This situation arises due toincompleteness of the traces. This can be caused by insufficientexecutions (b1? x2 is possible but was not observed) or insufficientinformation in the trace (the m events that precede x2 differ fromthose that precede y2, in a way that is not recorded in the trace). Ineither case, the b1 6? x2 invariant is an undesirable false positive.CSight detects and omits invariants of the above form by checkingif the paths in the concrete FSM satisfy the mined invariants. We callthis process invariant validation (step A in Figure 2). Figure 3(b)shows the traces from Figure 3(a) with intermediate anonymousstates ? note that the state of process p2 after executing ?m isidentical across all three traces. These anonymous local processstates, along with concrete channel contents, are used to derive theconcrete FSM (Figure 3(c)) corresponding to the above executionDAGs. This concrete FSM is checked for paths that violate themined invariants, and these invariants are reported to the user andautomatically omitted from the set of CSight steps that follow.The output of this process is a set of valid invariants. Figure 4details the ValidateInvariants procedure.It is worth noting that invariant validation is necessary becauseof how CSight generates anonymous local process states: the localprocess state is completely determined by the set of executed localevents (the assumption mentioned in Section 3.3). If a process statewas determined not by just the local process events, but by the globalhistory of events at all processes in the system then we would notneed to perform invariant validation (all invariants would be satisfiedin the concrete FSM). However, such an approach would lead tostate explosion, as states would be differentiated to a very fine degree.An advantage of our design choice is that it mitigates the much moredifficult problem of state explosion.One concrete method by which an invalid invariant can be madevalid is by refining the existing event types in the log based on remoteevents. For example, in the traces in Figure 4(a), if message m wasdifferentiated into mac, ma, and mb then all of the mined invariantswould be valid.3.4 Abstracting a concrete FSMA concrete FSM is not concise ? it is a DAG whose longestpath is as long as the longest execution ? and it is not sufficientlyabstract. CSight generates a more concise abstract FSM model fromthe concrete FSM, using a process we call state abstraction (step4 in Figure 2). As a reminder, we use the symbol ? to denotes0,t0[ ]Stitchingssatisfying all invariantsStitching invalidating b1 6! x2(c) Concrete FSMc1 ! y2Stitching invalidating p1 p2p1 p2p1 p2?mx?myTrace 2 Trace 3?myTrace 1s0s1aa!mb!ma!mcp1 p2p1 p2s2cs3!mt0t1?mt2ys0s1as4!mt0t1?mt3xp1 p2s0s5bs6!mt0t1?mt2yTrace 2 Trace 3Trace 1(a) System traces(b) Traces with anonymous statess3,t2[ ]s1,t0[ ]s4,t0[m]as5,t0[ ]b!ms4,t1[ ]?ms4,t3[ ]xs6,t0[m]!ms6,t1[ ]?ms6,t2[ ]ys4,t2[ ]ys6,t3[ ]s2,t0[ ]s3,t0[m]s3,t1[ ]?myc!ms3,t3[ ]xxFigure 3: (a) System traces parsed from an input log with threetraces, generated by two processes: p1 and p2. All messagesflow from p1 to p2. Note that the system traces satisfy the in-variants c1 ? y2 and b1 6? x2. (b) The traces from (a) withadded per-process anonymous states. Note the reuse of statess0, s1, t0, t1, and t2. (c) Concrete FSM for the system tracesin (a). The middle box highlights a stitching that satisfies all ofthe mined invariants. The two shaded boxes highlight stitchingsthat invalidate the c1 ? y2 and the b1 6? x2 invariants. Duringinvariant validation, CSight model-checks the concrete FSM toidentify and omit these invalid invariants from the refinementprocess.variables that represent abstract objects in the modeling process andto differentiate these from concrete features of the input log.The concrete FSM FL accepts all possible linearized sequencesof event instances from executions in a log L. Let P? represent apartitioning of states in FL, that is a partitioning of Q?M (we5University of British Columbia Technical Report1 function ValidateInvariants(FL, Invs):2 let Invs? = Invs3 foreach complete path p in FL:4 foreach inv ? Invs:5 if (p violates inv):6 Invs? = Invs? \{inv}7 return Invs?Figure 4: ValidateInvariants model-checks the mined in-variants Invs in the concrete FSM FL and returns a subset ofinvariants, Invs? ? Invs, that are valid for FL. This procedurecorresponds to step A in Figure 2.use bold font to denote a set of sets)3. CSight?s aim can now bedefined as deriving an abstract FSM (AFSM) A?L(P?) whose statesare partitions in P? and that accepts sequences of events ? rather thanevent instances. Transitions between states (partitions) in A?L(P?) aregenerated through existential abstraction: there is a transition fromP?1 ? P? to P?2 ? P? on event type e? iff there are concrete states s1 ands2 such that s1 ? P?1,s2 ? P?2, and there is a transition from s1 to s2on some event instance e, corresponding to e?. Note that A?L(P?), likethe concrete FSM FL, accepts all of the linearized event instancesequences. We now formally define A?L(P?).Definition 12 (Abstract FSM (AFSM)). For a log L, let FL = ?S, sI ,E, ?, ST ? be the concrete FSM for L, and let P? be a partitioning ofS. Then, an abstraction of FL, or an abstract FSM of L, is an FSMA?L(P?) = ?P?, P?I , E?, ??, P?T ?, where? Initial states in A?L(P?) are those partitions that contain theinitial concrete state: P?I = {P? ? P? | sI ? P?}.? Transitions in A?L(P?) are event types that correspond to eventinstances in the concrete FSM: E? = {e? | ?e,e ? E}.? A transition between two partitions in A?L(P?) exists if and onlyif there is a corresponding concrete transition between sometwo concrete states in the two partitions: ??(P?, e?) = P?? ???q ? P?,q? ? P??,e ? E,?(q,e) = q?.? Terminal states in A?L(P?) are those partitions that contain aterminal concrete state: P?T = {P? ? P? | ST ? P? 6= /0}.An important feature of an AFSM is that it generalizes observedsystem states. A partition contains a finite number of observed sys-tem states, but through loops with transitions that modify channelstate, an AFSM can generate arbitrarily long channel contents, lead-ing to an arbitrarily large number of system states. We may not haveobserved these system states, but an AFSM model generalizes topredict that they are feasible.CSight uses a first-k-in-channels partitioning strategy for gen-erating an initial AFSM for a concrete FSM. This partitioning as-signs two system states to the same partition if and only if thefirst-k message sequences in the channel states of the two statesare identical. For example, suppose a system has two channels,c12 and c21, and there are three concrete states: s1, s2, and s3.Let s.channels denote the channel contents for state s and supposethat s1.channels = {c12 : [],c21 : [m]}, s2.channels = {c12 : [],c21 :[m,m]}, and s3.channels = {c12 : [l],c21 : [m]}. Then the first-1 con-tents of s1 and s2 are {c12 : [],c21 : [m]} (the second m in s2.c21 isnot included), while the first-1 contents of s3 are {c12 : [l],c21 : [m]}.Therefore, in a first-1 partitioning strategy, s1 and s2 would map tothe same partition, while s3 would map to a different partition.3We use ? for P? even though it is a partitioning of states in theconcrete FSM because the partitions in this set represent states inthe abstract FSM. This set links the concrete and abstract FSMs.1 function AFSMtoCFSM(AFSM A?):2 A? = ?P?,PI ,E,?,PT ?3 foreach i ? [1, . . . ,h]:4 // A?i is A? with non pid i events replaced with ?.5 let A?i = (P?,PI ,Ei,?i,PT ), whereEi = {ei ? E}, and?i(q,?) = q? ?? ?(q,e j) = q?, i 6= j?i(q,ei) = q? ?? ?(q,ei) = q?6 let Fi = eliminate?? (A?i)7 CFSM C?A? = ?Fi?hi=18 return C?A?Figure 5: AFSMtoCFSM translates an AFSM A? into a CFSMC?A? ; eliminate-? performs standard ? transition elimina-tion [35]. This procedure corresponds to step B in Figure 2.RefinementA?0 A?1 A?final? ? ?100%15%10%Abstract FSMs:Invariants satisfied: ? ? ?Model Checker? ? ?CFSMmodelFigure 6: A high-level summary of the CSight refinement pro-cess (steps 5 and 6 in Figure 2). CSight starts with an ini-tial abstract FSM, A?0, which is refined to a final abstract FSM,A?final. This final abstract FSM satisfies all of the mined invari-ants, and is converted into a CFSM that is returned as output.Definition 13 (First-k partitioning). Let S be a set of system states.P?k is a first-k partitioning of S if ?P? ? P?k, ?q,w?,?q?,w?? ? P? ???ci j,?g ? [1,k],wi j[g] = w?i j[g].Finally, a CFSM is a set of per-process FSMs that communicateover FIFO channels. Channels are reliable, unidirectional, and havea single sender and receiver process. The alphabet of a processFSM includes process-local events, and inter-process communica-tion (message send and message receive) events. We use the notationc!m for a send event of message m on channel c, and the notationc?m for a receive event of message m on channel c. Each channelhas its own set of valid messages.Definition 14 (Communicating FSM (CFSM)). A CFSM of h pro-cesses is a tuple of h process FSMs, ?F1,F2, . . . ,Fh?, and a set mes-sage sets M to indicate the message types that can be sent by oneprocess and received at another. For each process i, its FSM isF?i = ?Q?i, I?i, E?i, ??i, T?i? and the message set Mi contains message typesthat can be sent from process i and received by some other process j,for all j. That is, Mi = ? jMi j, where ci j?m ? E? j ?? ?i,m ?Mi jand ci j!m ? E?i ?? ? j,m ?Mi j.3.5 Model-checking and refining an AFSMFigure 6 overviews, at a high level, the CSight model-checkingand refinement process and Figure 7 lists an outline of the completeCSight algorithm. CSight uses the McScM [33] model checker tocheck if an invariant holds in the AFSM (step 5 in Figure 2). AsMcScM model-checks CFSMs and not AFSMs, to use McScM,CSight converts an AFSM into a CFSM. Further, as McScM reasonsabout state (un-)reachability, CSight encodes a temporal invariantin terms of states that can only be reached if the sequence of exe-cuted events violates the invariant. This encoding is described in6University of British Columbia Technical Report1 function CSight(Log L,k):2 let Invs = ValidateInvariants(MineInvariants(L))3 let FL = concrete FSM for L4 let A? = AFSM for FL with first-k partitioning P?k5 let C?A? = AFSMtoCFSM(A?)6 let Invs = ValidateInvariants(MineInvariants(L))7 foreach Inv ? Invs:8 while (C?A? violates Inv): // Call to model checker.9 let p? = counter-example path for Inv in C?A?10 // Translate events path p? in C?A? into S?, a list of11 // sets of paths in A? , |S?|= h12 let S? = CFSMPathToAFSMPath(p, C?A? )13 A? = Refine(A? , S?)14 C?A? = AFSMtoCFSM(A?)15 return C?A?Figure 7: An outline of the CSight algorithm. Figure 5 de-tails the AFSMtoCFSM procedure and Figure 9 details theCFSMPathToAFSMPath procedure. MineInvariants is de-scribed in [9] and Figure 4 details ValidateInvariants.The CSight implementation also handles model checker time-outs.a? b?baP?1 P?2 P?3Figure 8: An AFSM path, [a?, b?], that can be eliminated by refin-ing (splitting) the abstract state P?2, separating the two concretestates that generate the abstract path.Appendix A.An AFSM is an abstraction of concrete FSM. CSight?s goal isto construct a communicating FSM (CFSM). An AFSM can bethought of as a cross product of the per-process FSM, and the CFSMcan therefore be reconstructed from the AFSM. Figure 5 details theAFSMtoCFSM procedure for converting an AFSM into a CFSM to useMcScM (step B in Figure 2). This procedure is also used to derivethe final CFSM output in CSight.Model-checking an invariant produces one of three cases:(1) The invariant holds in the model. There is nothing more forCSight to do for this invariant.(2) The invariant does not hold and McScM finds and reports acounter-example CFSM execution. A CFSM execution is a sequenceof events that abides by CFSM semantics (e.g., a process can onlyreceive a message if that message is at the head of the channel).A counter-example CFSM execution is a sequence of events thatviolates the invariant. In this case, CSight refines the AFSM toeliminate the counter-example (Section 3.5.1) and then re-model-checks this invariant in case another counter-example for it exists.(3) McScM fails to terminate within a user-defined threshold (de-faulted to 5 minutes). CSight then attempts to check a differentinvariant first before coming back to the invariant that timed out.4If CSight resolves all invariants, it outputs the model. But, ifMcScM repeatedly times out on every invariant, then CSight termi-nates and does not output a model. We did not observe the latter4After refining the model to satisfy one invariant, a different, previ-ously difficult-to-check invariant may become trivial to model-check.1 function CFSMPathToAFSMPath(p?, C?A? ):2 // p? is a complete events path in CFSM C?A?3 let AFSM A? = ?P?, P?I , E?, ??, P?T ?4 foreach i ? [1, . . . ,h]:5 let Ei = event types generated by process i6 let S?i = {s? | s? an events path in A? from P?1 to P?k,pi(s?,Ei) = pi(p?,Ei), P?1 ? P?I ,??q,w? ? P?k,qi terminal,and? j,w[ ji] = ?}7 return [S?1, . . . , S?h]Figure 9: CFSMPathToAFSMPath translates an events path p?in a CFSM C?A? into S?, a list of sets of paths in A? . Each set in S?maps to one event in p?.case in our experiments.3.5.1 Refining an AFSM to satisfy invariantsCSight uses counter-example guided abstraction refinement (CE-GAR) [16] to eliminate a counter-example for an invariant (step6 in Figure 2). Figure 8 illustrates how the concrete states from thelog may generate a counter-example partitions path in the AFSM.The McScM-generated invariant counter-example is a CFSMexecution. Refining CFSM states to eliminate a counter-exampleis challenging because a single state in the concrete FSM can mapto multiple states in the CFSM. Instead, CSight performs partitionrefinement in the AFSM. It does this by mapping the McScM-generated CFSM counter-example into an AFSM counter-example.Figure 9 describes this translation.Note that the AFSM counter-example is a list of sets of AFSMpaths, one set of AFSM paths for each process in the system. This isbecause the process-specific event subsequence of a CFSM executionmaps to potentially multiple paths in the AFSM (due to the CFSMconstruction based on ?-transitions in Figure 5).Once the AFSM counter-example is generated, CSight eliminatesthe CFSM counter-example by transforming the AFSM into a moreconcrete (or less abstract) AFSM. It does so using partition refine-ment (Refine in Figure 10). Given an AFSM counter-example,Refine identifies the set of partitions that stitch concrete observa-tions, as in partition P?2 in Figure 8. It then refines all partitionsin a set that is smallest across all processes and returns the refinedAFSM.A refined AFSM is more concrete ? closer to the the concreteFSM. The refined AFSM has more partition states, and each partitionstate contains fewer concrete system states.Definition 15 (AFSM Refinement). An AFSM A?L(Q?) is a refine-ment of AFSM A?L(P?) if ?Q? ? Q?,?P? ? P?, Q?? P?, and?Q??Q?Q? =?P??P?P?Next, in Section 4, we prove three key properties of the CSightprocess.4. FORMAL ANALYSISWe now use the formalisms defined in the previous section toprove three results about the CSight model inference process: (1) theinferred model fits the input log, (2) the inferred model satisfiesmined invariants, and (3) refinement always makes progress.We begin with an observation: The concrete FSM FL satisfiesall mined, validated invariants. This is true by construction inValidateInvariants in Figure 4.7University of British Columbia Technical Reports1s2SC!syn-acks3CS?acks4SC!ackCS?syn(a) TCP server (b) TCP clients0SC!fins5s6SC!acks9CS?ackCS?fins10SC!acks7CS?fins11CS?acks8SC!acks12SC!finCS?ackc1c2SC?syn-ackc3CS!ackc4SC?ackCS!sync0c5c6CS!ackc10SC?finCS!finc11SC?ackc7SC?ackc12CS!ackc9CS!finc13SC?finCS!ackc8SC?ackSCCSFigure 11: The inferred state machines for the (a) TCP server and (b) TCP client. The server communicates with the client over the SCchannel, and the client communicates with the server over the CS channel. Shaded states represent the connection-establishedstates.1 function Refine(AFSM A? , AFSM Event Paths S?):2 let ?P?, P?I , E?, ??, P?T ?= A?3 // min is an index into S?, denoting a set of process paths4 // requiring the fewest number of refinements to eliminate.5 let min = 06 foreach i ? [1, . . . ,h]:7 foreach s? ? S?[i]:8 let ??= state sequence for s? in A?9 // Find stitching partitions, e.g., P?2 in Fig. 8, by10 // traversing ?? and recording partitions that can be11 // refined to eliminate ?? from A? .12 let S?titchs? = {P? | P? is a stitching state in ??}13 if S?titchs? is empty:14 next i15 // Set of stitching partitions shared by those paths in A?16 // that correspond to strings in S?[i]17 let S?titchi = ?s??S?[i]S?titchs?18 // min is an index that tracks the smallest S?titchi19 if min = 0 or???S?titchi???<???S?titchmin???:20 min = i21 let P?? = P? with all partitions in S?titchmin refined to22 eliminate all paths in S?[min]23 // Derive P??I , ???, and P??T from P?? as in Def. 12.24 let AFSM A? ? = ?P??, P??I , E?, ???, P??T ?25 return A? ?Figure 10: Refine removes an invariant counter-examplefrom an AFSM. It refines one of the sets of process paths inS?, selecting the one that require the fewest refinements. Notethat S? is a list of sets of paths, one set per process.Observation 1 (Concrete FSM satisfies mined, validated invariants).Let L be a log, and let Invs be the set of invariants that are valid inFL. Then, ?Inv ? Invs, s ? Lang(FL), s satisfies Inv.Theorem 1 (Inferred model fits the input log). For all logs L andintegers k, CSight(L,k) returns a CFSM model that accepts alltraces in L.Proof of Theorem 1. Let C?A? be the model returned by CSight(L,k).By construction, A? accepts all of the traces in C?A? . Furthermore, A?is an abstract FSM for FL ? the concrete FSM of log L.All abstract FSMs must at least accept the traces in the concreteFSM. Therefore, A? accepts all of the traces in FL. Since, FL acceptsall of the traces in L by construction, A? accepts the traces as well,and therefore C?A? must also accept them.A key property of the Refine procedure in Figure 10 is that iteliminates a counter-example path from a CFSM. We prove thisnext.Theorem 2 (Refinement progress: Refinement eliminates coun-ter-examples). Let p? be a CFSM counter-example path for invariantInv in C?A? , let S? = CFSMPathToAFSMPath(p?, C?A? ), and let A? ? =Refine(A? , S?). Then, p? is not a counter-example to Inv in C?A? ? . Thatis, p? is not a valid execution of C?A? ? .Proof of Theorem 2. Proof by contradiction. Assume that p? is asequence of events that is a valid execution of C?A? ? and that p? violatesInv. Consider an invocation of Refine(A? , S?). Refine (Figure 10)uses a non-zero min value to compute P?? on line 19 and returns aCFSM C?A? ? = ?F?i?hi=1.For p? to be a valid execution in C?A? ? , the sub-sequence of process ievents p?i, p?i = pi(p?, E?i), must be a valid execution in F?i, for all i. Theprocedure AFSMtoCFSM (line 14 in Figure 7), detailed in Figure 5,constructs F?i to accept p?i iff there is a complete path s? in A? ?, suchthat p?i = pi(s?, E?i). However, any such s? must also be in ?Stitchmin,which is used to compute P?? in Figure 10. Therefore, after refining?Stitchmin, s? can no longer be a valid path in A? ?. Contradiction.Now, we prove that the CSight procedure in Figure 7 returns aCFSM model that satisfies all of the mined, validated invariants.Theorem 3 (Inferred model satisfies mined invariants.). For a givenlog L, if CSight outputs a CFSM model then this model satisfiesall of the mined, validated event invariants from FL.Proof of Theorem 3. For a log L with a total of n event instances,CSight can refine the initial abstract FSM for L, AL(P?k), at mostn?1 times. This is because after n?1 refinements, each partitionin the abstract FSM must map to exactly one concrete state, and asingleton partition cannot be refined further.Let A? be the abstract FSM after n? 1 refinements of AL(P?k).Because A? maps each event instance to a unique partition, it isindistinguishable from FL, the concrete FSM it abstracts. Therefore,Lang(A?) = Lang(FL). By Observation 1, FL satisfies all validatedinvariants, therefore so does A? .Since CSight does not terminate until all the validated invariantsare satisfied in the abstract FSM, it either returns A? after n? 1refinements, or it returns a smaller (and more abstract) A? ?. In bothcases, the returned AFSM satisfies all of the validated invariants.8University of British Columbia Technical Report5. LOGGING VECTOR TIMESTAMPSCSight requires its input logs to be annotated with vector times-tamps. Vector time is a logical clock mechanism to partially orderevents in a concurrent system [25, 46]. While some systems use thismechanism internally and could include vector timestamps in theirlogs, many do not. To ease using CSight with concurrent systems,we have built ShiVector5, a tool to automatically compute and insertvector timestamps into Java system logs. ShiVector does not requiremodification of the system?s source code, only recompilation withthe ShiVector library added to the classpath.ShiVector uses aspects (via AspectJ) to intercept socket-basednetwork communication to piggy-back process vector timestamps ontop of existing application-level protocols.6 ShiVector also interceptslogging events to augment each logged event with a vector timestampat the time that the event was logged. Thus, ShiVector automaticallytracks and injects vector timestamps into concurrent systems? logs.6. EXPERIMENTAL EVALUATIONThe CSight prototype uses McScM for model checking andgraphviz for model visualization. In all experiments we used avalue of k = 1 for the first-k-in-channels partitioning strategy. Weevaluated CSight on three sets of logs, produced by: a simulatorof the stop-and-wait protocol; the TCP stack of OS X; and Volde-mort [57], an open-source project that implements a distributed hashtable [22] and is used in data centers at companies like LinkedIn. Wealso carried out a user study with a group of undergraduate studentsto evaluate the efficacy of CFSM models in finding bugs.6.1 Stop-and-wait protocolWe applied CSight to traces from a simulator of the stop-and-wait protocol described in Section 2. We derived a diverse setof traces by varying message delays to produce different messageinterleavings. CSight mined a total of 66 valid invariants. Themodel CSight derived (Figure 1(b)) is identical to the true model ofthe stop-and-wait protocol. This experiment was a sanity check toverify that CSight performed as expected on this well-understoodprotocol, when faced with concurrency-induced non-determinism ininterleavings.6.2 TCPThe TCP protocol uses a three-phase opening handshake to estab-lish a bi-directional communication channel between two end-points.It tears down and cleans up the connection using a four-phase closinghandshake. The TCP state machine is complicated by the fact thatpacket delays and packet losses cause the end-points to timeout andre-transmit certain packets, which may in turn induce new messages.Our goal was to model common-case TCP behavior, so we did notexplore these protocol corner cases.We used netcat and dummynet [52] to generate TCP packet flow.We captured packets using tcpdump and then semi-manually anno-tated the log to include vector timestamps. A subset of the tracesthat involve only the the opening and closing TCP handshakes werefed into CSight.For the captured TCP log, CSight identified 149 valid invari-ants, some of which are not true of the complete protocol (e.g.,because the input traces did not contain certain packet retrans-missions). The CSight-derived CFSM model is shown in Fig-ure 11. The shaded states s4 and c4 represent the server and clientconnection-established states, which is reached when the two5https://bitbucket.org/bestchai/shivector/6We plan to extend ShiVector to interpose on other kinds of inter-process communication.end-points have successfully set up the bi-directional channel. Tran-sitions up to these two states model the opening handshake, whiletransitions after these states model the closing handshake. The clos-ing handshake is split into a server-initiated tear-down sequence(middle row of states) and a client-initiated tear-down (bottom-mostrow of states).The derived model is accurate except for the self-loop on states4 in Figure 11(a). This loop appears because s4 represents boththe connection-established state and the state after the serverhas initiated the closing handshake. This loop appears to contradictthe SC!fin 6? SC!fin invariant, which is mined by CSight and isvalid. However, the model checker only considers counter-examplesthat terminate. Note that if the loop at s4 is traversed twice thenthe client will be unable to consume both server fin packet copiesand will enter an unspecified reception error state and therefore notterminate. Such unspecified reception states are typically undesir-able, as they can be confusing. The McScM model checker can beused to detect these states and further refinement will eliminate them.Implementing this elimination remains as part of future work.Figure 11 also illustrates a key feature of CSight ? the userdecides (by specifying a set of line-matching regular expressions)what information in the log is relevant to the modeling task. Thus,for each use of CSight, the user decides on the trade-off betweencomprehensibility of the model (e.g., model size) and the amount ofinformation lost/retained by the modeling process. For example, theTCP model in Figure 11 is simple to understand, but it omits TCPsequence numbers, data packets, and other details that were presentin the input log.6.3 Voldemort distributed hash tableVoldemort [57] implements a distributed hash table with a clientAPI that has two main methods. put(k,v) associates the value vwith the key k, and get(k) retrieves the current value associatedwith the key k. Voldemort is a distributed system as it providesscalability by partitioning the key space across multiple machinesand achieves fault tolerance by replicating keys and values acrossmultiple machines.The Voldemort project has an extensive test suite. We used asubset of integration tests that exercise the synchronous replicationprotocol to generate a log of replication messages in a system withone client and two replicas. We logged messages generated by clientcalls to the synchronized versions of put and get and captured justthe messages between the client and the two replicas.7 Since Volde-mort does not implement vector timestamps, we used ShiVector(described in Section 5) to produce a vector timestamped log.CSight mined 112 valid invariants and generated the model in Fig-ure 12. This model contains a client FSM and two replica FSMs. Asexpected, the replica FSMs are identical. Synchronized Voldemortoperations are serialized in a specific order, so the flow of messagesfor put as well as for get is identical ? the client first executes theoperation at replica 1 and then at replica 2.We manually inspected Voldemort?s replication code to confirmthat the model in Figure 12 is accurate. This model provides ahigh-level overview of how replication messages flow in the system.However, as with the TCP protocol, the model also abstract awayand omits numerous details, such as what happens when replicas fail.One feature of CSight is that it allows a developer to focus on thoseaspects of system behavior that are important to them. For example,by not exercising irrelevant system behavior in the first place, or byrunning CSight with regular expressions that ignore certain loggedbehavior in the log parsing stage of the process.7The replicas also communicate with each other to maintain keyavailability, but we excluded this communication from the log.9University of British Columbia Technical Report(a) Voldemort Replica 1 (b) Voldemort clientc1c2R1-C?put-rec3C-R2!putR2-C?put-reC-R1!putc0R1-CC-R1c4c5R1-C?get-rec6C-R2!getR2-C?get-reC-R1!getr1r0r2R1-C!get-reC-R1?get C-R1?putR1-C!put-re(c) Voldemort Replica 2s1s0s2R2-C!get-reC-R2?get C-R2?putR2-C!put-reR2-CC-R2Figure 12: (a,c) replica models and (b) client model for a three-node Voldemort cluster. ?re? stands for ?response?.6.4 User studyTo determine whether CFSM models are useful in finding bugs,we designed and ran a user study. It is typical for developers tointerpret distributed executions in the form of time-space diagrams.Therefore, we compared the efficacy of CFSM models against time-space diagrams in bug-finding tasks. Throughout the study and inthe oral feedback session that followed, participants were not toldthe purpose of the study.The study focused on two concurrent systems: the stop-and-waitprotocol and Voldemort. We introduced a bug into each system ?the sender in the stop-and-wait protocol failed to re-transmit packetson timeout, and the synchronized Voldemort client sent requests toall replicas concurrently, instead of blocking on acknowledgmentfrom each replica. In our experience, both bugs are representativeof real bugs faced by distributed system developers. The root causeof the stop-and-wait bug is not taking the right action on an event;the root cause of the Voldemort bug is performing an action atthe wrong time. For each buggy system, we generated: (1) a setof representative time-space execution diagrams (8 for the stop-and-wait protocol and 6 for Voldemort), and (2) a CFSM model.Overall, we created four artifacts for the study ? tspace-saw andcfsm-saw (time-space diagrams and CFSM model of the buggy stop-and-wait protocol), and tspace-vol and cfsm-vol (artifacts for buggyVoldemort).The study consisted of an in-class, web-based assignment8 inan undergraduate Introduction to Software Engineering class atthe University of Massachusetts, Amherst. The 39 students whocompleted the assignment had, on average, 4.2 years of programmingexperience, and 76% of the students had never taken a networkscourse.We considered two factors: the model factor (time-space diagramsvs. CFSM models), and the task factor (stop-and-wait vs. Voldemort).To account for learning effects, we used a within-participants mixeddesign across all 39 participants. We randomly assigned each studentto one of four possible study sequences: ?tspace-saw, cfsm-vol?,?tspace-vol, cfsm-saw?, ?cfsm-saw, tspace-vol?, ?cfsm-vol, tspace-saw?. We considered the two bugs to be independent ? findingone bug in one artifact does not help in finding the other bug in adifferent artifact.Each task consisted of two steps:1. Each student completed a mini-tutorial on the appropriatediagram (time-space and CFSM, respectively). To verify their under-standing, at the end of each tutorial, each student had to answer twobasic questions correctly. They could resubmit their answers untilboth were correct.92. Each student was given a correct, written description of the8Available online: http://www.bestchai.net/papers/icse2014_csight_eval/9The average time to complete either tutorial was 5 minutes. Stu-dents made, on average, 1.8 attempts to answer the tutorial questionscorrectly. These measures were similar when the time-space and theCFSM tutorials were considered separately.system in English, along with either a set of time-space diagrams ora CFSM model. The student was asked to respond to a single open-ended question. For time-space diagrams (tspace-saw, tspace-vol)we asked: ?List all of the time-space diagrams above that you thinkdo not conform to the description of the system above. What madeyou choose these diagrams?? For the CFSM models (cfsm-saw,cfsm-vol) we asked: ?How does the observed model differ from theintended system description??6.4.1 ResultsStudents found bugs approximately as well with the CFSM modelas they did with the time-space diagrams that are developers? currentpreferred visualization. For the stop-and-wait protocol, studentswho were shown the CFSM were 72% successful in getting the rightanswer, while those who were shown the 8 time-space diagramswere 61% successful. For Voldemort, the success rate with theCFSM was 72% and the success rate with the 6 time-space diagramswas 86%.Students found CFSMs just as useful (for completing the task) asthe small collection of time-space diagrams. These results indicatethat CFSM models can be used to effectively find bugs. Moreover,in practice, developers have to inspect neither 8 nor 6 executionsof the system, but hundreds or thousands. The task of manuallyfinding the anomalous time-space diagrams would be infeasible.The CFSM models, though, will remain roughly of the same sizeand complexity regardless of the number of executions. Therefore,we believe that our results on the utility of CFSMs over time-spacediagrams are conservative, and CFSMs would perform even betterin practice. Our study finds that they already perform as well astime-space diagrams tightly focused on the buggy behavior.Students? oral feedback on the assignment reveals why manyof them preferred the CFSM model. According to many students,time-space diagrams were difficult to follow, especially for longexecutions:?I found the time-space diagrams confusing. It was hard to tell whatwas happening when. The [CFSM] models were simpler and made itclear what state a system was in and I could keep track of that state.? ?student 1?Time-space diagrams were easy to follow for small time segments. Forlonger time segments, you got lost. The other [CFSM] models werebetter for longer time.? ? student 2Students could compare a CFSM model to their own mentalmodel, or the model they would draw after reading the systemdescription:?I read the description and tried to recreate the [CFSM] model myselffirst. Then, I compared what I drew to the given diagram and foundmistakes. Understanding those mistakes helped me understand thesystem a lot better.? ? student 3Finally, students mentioned that CFSM models were easier to usebecause they did not explicitly model time:?For the [CFSM] models, I assumed no time delay in the network mes-sages. It was harder to do in the time-space diagrams because you10University of British Columbia Technical Reportcannot ignore the delay there. In CFSMs, you can ignore time at first,and then allow for it.? ? student 47. RELATED WORKCFSMs can be inferred from manually-labeled message sequencecharts [13]. In contrast, CSight automates the inference by relying onmined invariants. CFSMs inferred from executions can demonstratesystem properties, such as absence of deadlocks and unspecifiedreceptions [15, 56]. We found that CFSMs also provide a conciserepresentation of complex concurrent system logs. Another poten-tially useful inferred model of concurrent systems is class-levelspecifications in the form of symbolic message sequence charts [39].CSight can use these to merge identical process FSMs, such as thereplica models in Figure 12(a,c).In addition to inferring models, concurrent system logs can beused to detect anomalies [36, 45, 58], identify performance bugs [53,54], and mine temporal system properties [9, 17, 60]. Our focus ison concurrency and on extracting a model that can aid understand-ing of more general system behavior. Our own work on miningtemporal invariants of concurrent systems [9] is complementary toCSight, which needs to mine such invariants from execution logs.Probabilistic analysis of logs (measuring number, frequency, andregularity of event occurrences) can help discover concurrent systemexecution patterns, which can help the developer better understandand improve the design of concurrent systems [17]. Logs can alsobe used to mine message sequence charts [38], which describe theprocess interactions within a concurrent system.The problem of finding a short sequence of refinements to producea small abstract FSM that satisfies all of the valid invariants is NP-hard [16], and CSight?s design finds an approximate solution. Forlogs of serial systems (totally ordered logs), the problem of automatainference from positive examples of executions is computable [12]but NP-complete [30, 5], and the FSA cannot be approximated byany polynomial-time algorithm [49]. Unlike CSight, prior work onmodel inference from totally ordered logs either excluded concur-rency or captured a particular interleaving of concurrent events [2,10, 7, 11, 18, 28, 44, 47, 50]. Tomte [1] and Synoptic [10] take asimilar approach to CSight, using CEGAR [16] to refine models.Synoptic mines temporal invariants and infers FSM models froma log of sequential executions, while Tomte infers scalarset Mealymachines. CSight mirrors Synoptic?s inference procedure but usesa different modeling formalism and algorithms, and works for con-current systems that log concurrency as a partial order. We believethat modeling concurrency explicitly is crucial to understanding aconcurrent system?s behavior.System models can also be inferred from developer-written specifi-cations [4, 19, 21, 29, 31, 37]. Compared to CSight, these techniquesrequire significantly more manual effort from the developer and arenot suitable for legacy systems.CSight relies on the McScM model checker [33, 32] for check-ing the validity of invariants in a CFSM model. A key property ofMcScM is that it verifies models with no a priori bound on chan-nel size. However, this means that McScM is not guaranteed toterminate.Currently, the scalability bottleneck in CSight is the McScMmodel checker [33, 32], whose runtime depends on model size andmodel complexity. To improve scalability, we plan to use the moreefficient Spin model checker [34], though Spin is less precise onstates with channels that contain many messages.Aguilera et al. pioneered a popular black-box approach to debug-ging distributed systems [3]. In this approach, multiple observationsof a distributed executions are used to infer causality, dependencies,and other important characteristics of the system. This approach wassomewhat relaxed in later work to produce more informative andapplication specific results in Magpie [6] and later in X-Trace [26].The focus of this prior work is on data, or request flow. In contrast,CSight generates CFSM models that model the partially orderedsequence of events executed by a concurrent system.Runtime checking of concurrent systems can also aid debuggingby monitoring user-defined system invariants or properties as thesystem executes. If a property is violated, reporting the incidentto the user might identify a bug [51, 27, 42, 20]. Alternatively,the system could self-adapt by steering the system away from theviolation [41, 23, 59, 48]. CSight could be used to first infer a model,and then check, at runtime, that the executions adhere to the model.8. DISCUSSIONThe vector timestamps tracked and logged by the ShiVector library(Section 5) may adversely affect system performance. However, asis typical for production systems, ShiVector can be enabled for asmall fraction of the requests. As long as the collected traces arerepresentative of typical system behavior, CSight?s model, too, willbe representative.CSight?s model construction cannot directly eliminate two kindsof error states [14] ? unspecified reception and deadlock. Unspeci-fied reception occurs when a process enters a state with a messagem at the head of its channel, but has no reachable future state thatreceives m. A deadlocked system state occurs when no process cansend a message and at least one process cannot reach a terminatingstate. Currently, CSight does not check if these error states are reach-able in the final model. It is possible to extend CSight with such acheck, for example, by using the McScM model checker, but thecheck would be computationally expensive.CSight mines three types of invariants and ensures that the finalmodel satisfies the valid invariants. For the invariants to be accurate,the log must have executions representative of all possible executionsof the modeled system. While we found these invariant types tobe sufficient to infer interesting models in practice, more extensiveinvariants can lead to more expressive models.CSight works for system traces that satisfy certain communicationconstraints (Def. 1). For example, CSight cannot model uncleantermination and assumes that each execution terminates with emptychannels. However, in practice, system logs may not satisfy theseconstraints. In our future work we will extend CSight to handleterminal states with non-empty channels.9. CONCLUSIONConcurrent systems are hard to implement, debug, and verify.To help with these tasks, we developed CSight, a tool that usesa partially ordered log of events to infer a concise and accuratecommunicating finite state machine model of the system. CSight?saccuracy comes from its use of mined temporal properties that relateevents in the log. Our tool prototypes are available for download athttp://synoptic.googlecode.com/. By automatically mining a systemmodel from logs, CSight has the potential to ease system understand-ing, debugging, and maintenance tasks.Acknowledgments: We thank Jenny Abrahamson for her workon ShiVector and Roykrong Sukkerd for her contributions to theproject. This material is based upon work supported by the UnitedStates Air Force under Contract No. FA8750-12-C-0174 and byIARPA under Contract No. N66001-13-1-2006.11University of British Columbia Technical Report           qq'     q''q'''     aiq              q'cInv!aprei cInv!apostiaiFigure 13: An edge in a CFSM (top) that is transformed to trackai (bottom).APPENDIXA. CHECKING INVARIANTS WITH MCSCMModel checking an event invariant (Section 3.2) in a CFSM (Sec-tion 3.4) is a three step process. First, the CFSM must be augmentedwith synthetic ?tracking? transitions that record when an event typementioned by the invariant is executed by the model checker. Sec-ond, the negation of the invariant must be encoded as a set of ?badstates? of the CFSM. If the model checker finds an execution thatcan reach one of these states, it will emit the path, which is thecounter-example path for the invariant. Finally, the counter-examplepath must be post-processed before it can be used. We now describeeach of these steps in more detail.A.1 Preparing a CFSM for model checkingConsider an event invariant Inv, such that Inv = aiT b j , with T ?{?, 6?,?}. And, let C be a CFSM in which we want to check ifInv holds.We modify C to produce a new CFSM model, C ?, which will beused as input to the McScM model checker. We transform C intoC ? in two ways ? we convert local events into send events on asynthetic channel, and we add synthetic ?tracking events? to trackthe execution of ai and b j event types.Handling local events. The McScM model checker does notmodel local events (it only supports message send and receiveevents). We cannot omit local events from C ?, as we need to beable to unambiguously map an McScM counter-example path in C ?to a path in C . We therefore add a synthetic ?local events? chan-nel, clocal, and allow all processes to send messages on clocal. Thischannel is write-only ? processes never receive on clocal. Finally,each process local event type, ei, in C is translated into the eventtype clocal!ei in C ?. That is, we replace transitions of the form?i(q,ei) = q? with ?i(q,clocal!ei) = q?.Tracking events necessary for checking Inv. Given a CFSMand a set of ?bad states?, the McScM model checker checks if thereis an execution of the input CFSM that causes it to reach one ofthe bad states. A bad state is a kind of a system state (introducedin Section 3.3); it consists of a global process state and a globalchannel state. In McScM, the global channel state is expressed withregular expressions over channel contents.10To check Inv with McScM, we therefore need to generate a setof system states of C ?, each of which encodes a violation of Inv.To generate this encoding for Inv = aiT b j, we need to track whenai or b j occur. We do this with ?tracking events?, which track theoccurrence of ai and b j as send events to a synthetic Inv-specificchannel, cInv. This channel is used exclusively for checking Inv. Aswith the synthetic local events channel, all processes can send on cInv,and no process ever receives from this channel. Figure 13 illustratesthis tracking. More formally, if ?i is the process i FSM transitionfunction in C , then we track ai and b j event types as follows:1. Replace transitions of the form ?i(q,ai) = q? with:10Therefore, an McScM bad state is, in fact, a set of system states, asdefined in Section 3.3.? ?i(q,cInv!aprei ) = q??? ?i(q??,ai) = q???? ?i(q???,cInv!aposti ) = q?Where, q??,q??? are synthetic states with just the transitionsabove, and cInv!aprei and cInv!aposti are synthetic event typesthat maintain a record of when ai occurs as messages in cInv.2. Similarly, to keep track of b j, replace ? j(p,b j) = p? with:? ? j(p,cInv!bprej ) = p??? ? j(p??,b j) = p???? ? j(p???,cInv!bpostj ) = p?Note that the first transformation retains the event ai and thesecond transformation retains the event b j. This is necessary as wewant C ? to have identical behavior to C (ignoring the added trackingevents).To see why we need to augment both ai and b j with a pre anda post event, consider other strategies, with less overhead (fewersynthetic events). For example, assume that Inv = ai? b j and werepaced transitions of the form ?i(q,ai) = q? with ?i(q,cInv!aprei ) =q?? and ?i(q??,ai) = q?; and we replaced transitions of the form? j(p,b j) = p? with ? j(p,cInv!bprej ) = p?? and ? j(p??,b j) = p?.Then, consider the following contents of cInv: [aprei ,bprej ]. Thesetwo tracking events imply one of three executions:1. cInv!aprei ? ai? cInv!bprej ? b j Inv true2. cInv!aprei ? cInv!bprej ? ai? b j Inv true3. cInv!aprei ? cInv!bprej ? b j? ai Inv falseThe above example indicates that the simpler strategy results inambiguity ? given the contents of channel cInv we cannot tell ifInv satisfies the execution that produced the tracing events in cInv.Likewise, we can show that other strategies that do not include allof the four pre and post events lead to ambiguity.A.2 Expressing invariants as McScM bad statesTo model check Inv in C ? we need to construct bad states for Inv.Each such state needs to specify the states of all of the processes inthe system and describe the contents of all channels in C ? as regularexpressions. First, we overview the McScM regular expressionsyntax for denoting channel contents:? a . b concatenation of a and b? a?? zero or more a? a?+ one or more a? a | b either a or b? ( ) grouping? empty channelUsing the above syntax, we now define some basic patterns thatwe will reuse in specifying bad states.? ANY = (cInv!aprei | cInv!aposti | cInv!bprei | cInv!bposti )??Accepts an arbitrary sequence of tracking events.? A = (cInv!aprei | cInv!aposti )Implicitly accepts an ai via the corresponding tracking events.? B = (cInv!bprei | cInv!bposti )Implicitly accepts a b j via the corresponding tracking events.Now, let Acc be the set of all accepting global process states. Thatis, Acc = {a | a[i] accept state for process i}. We will now form badstates based on the kinds of invariant we are checking. Note that weonly specify the contents of cInv channel; the contents of all otherchannels are: (1) all system channels (non-invariant and non-local-event channels) are specified as being empty ? ? ?, and (2) the local12University of British Columbia Technical Reportevents channel is specified as containing an arbitrary sequence oflocal events.11Now, we can specify the set of bad states B for each invarianttype:? T =? B = {?a,ANY .A?+?|a ? Acc}? T =6? B = {?a,ANY .A .ANY .B .ANY?|a ? Acc}? T =? B = {?a,B?+ .ANY?|a ? Acc}A.3 Post-processing the counter-example pathThe invariant counter-example path returned by McScM for C ?must be post-processed to derive a valid counter-example for C . Thepath is transformed in two ways, that counter the transformationsdescribed in A.1: (1) remove all tracking events, and (2) replaceevents that are send events to the local channel, clocal, to be localevents in C .REFERENCES[1] F. Aarts, F. Heidarian, H. Kuppens, P. Olsen, and F. W. Vaan-drager. Automata learning through counterexample guidedabstraction refinement. In FM, 2012.[2] M. Acharya, T. Xie, J. Pei, and J. Xu. Mining API Patterns asPartial Orders from Source Code: From Usage Scenarios toSpecifications. In FSE, 2007.[3] M. K. Aguilera, J. C. Mogul, J. L. Wiener, P. Reynolds, andA. Muthitacharoen. Performance debugging for distributedsystems of black boxes. SIGOPS Oper. Syst. Rev., 37(5):74?89,Oct. 2003.[4] D. Alrajeh, J. Kramer, A. Russo, and S. Uchitel. LearningOperational Requirements from Goal Models. In ICSE, 2009.[5] D. Angluin. Finding Patterns Common to a Set of Strings.Journal of Computer and System Sciences, 21(1):46 ? 62, 1980.[6] P. Barham, A. Donnelly, R. Isaacs, and R. Mortier. Using mag-pie for request extraction and workload modelling. In NSDI,2004.[7] I. Beschastnikh, Y. Brun, J. Abrahamson, M. D. Ernst,and A. Krishnamurthy. Unifying FSM-Inference Algorithmsthrough Declarative Specification. In ICSE, 2013.[8] I. Beschastnikh, Y. Brun, M. D. Ernst, and A. Krishnamurthy.Inferring Models of Networked Systems from Logs of theirBehavior with CSight. In ICSE, 2014.[9] I. Beschastnikh, Y. Brun, M. D. Ernst, A. Krishnamurthy, andT. E. Anderson. Mining Temporal Invariants from PartiallyOrdered Logs. OSR, 45(3):39?46, Dec. 2011.[10] I. Beschastnikh, Y. Brun, S. Schneider, M. Sloan, and M. D.Ernst. Leveraging Existing Instrumentation to AutomaticallyInfer Invariant-Constrained Models. In ESEC/FSE, 2011.[11] A. W. Biermann and J. A. Feldman. On the synthesis of finite-state machines from samples of their behavior. IEEE Trans.Comput., 21(6):592?597, June 1972.[12] L. Blum and M. Blum. Toward a Mathematical Theory ofInductive Inference. Information and Control, 28(2):125 ? 155,1975.[13] B. Bollig, J. P. Katoen, C. Kern, and M. Leucker. LearningCommunicating Automata from MSCs. IEEE Trans. Softw.Eng., 36(3):390?408, May 2010.[14] D. Brand and P. Zafiropulo. On Communicating Finite-StateMachines. J. ACM, 30(2):323?342, Apr. 1983.[15] X. J. Chen and H. Ural. Construction of Deadlock-free De-signs of Communication Protocols from Observations. TheComputer Journal, 45(2):162?173, Jan. 2002.11This is (e1| . . . |en)??, where ei is a local event in C .[16] E. Clarke, O. Grumberg, S. Jha, Y. Lu, and H. Veith.Counterexample-guided Abstraction Refinement. In ComputerAided Verification, pages 154?169. Springer, 2000.[17] J. E. Cook, Z. Du, C. Liu, and A. L. Wolf. Discovering modelsof behavior for concurrent workflows. Computers in Industry,53(3):297?319, 2004.[18] J. E. Cook and A. L. Wolf. Discovering Models of SoftwareProcesses from Event-Based Data. TOSEM, 7(3):215?249,1998.[19] C. Damas, B. Lambeau, and A. van Lamsweerde. Scenarios,Goals, and State Machines: a Win-Win Partnership for ModelSynthesis. In FSE, 2006.[20] D. Dao, J. Albrecht, C. Killian, and A. Vahdat. Live debuggingof distributed systems. In ETAPS 2009, 2009.[21] G. de Caso, V. Braberman, D. Garbervetsky, and S. Uchitel.Validation of Contracts Using Enabledness Preserving FiniteState Abstractions. In ICSE, 2009.[22] G. DeCandia, D. Hastorun, M. Jampani, G. Kakulapati, A. Lak-shman, A. Pilchin, S. Sivasubramanian, P. Vosshall, and W. Vo-gels. Dynamo: Amazon?s Highly Available Key-Value Store.In SOSP, 2007.[23] B. Demsky, M. D. Ernst, P. J. Guo, S. McCamant, J. H. Perkins,and M. Rinard. Inference and enforcement of data structureconsistency specifications. In ISSTA, pages 233?243, July2006.[24] M. B. Dwyer, G. S. Avrunin, and J. C. Corbett. Patterns inProperty Specifications for Finite-State Verification. In ICSE,1999.[25] C. J. Fidge. Timestamps in Message-Passing Systems thatPreserve the Partial Ordering. In 11th Australian ComputerScience Conference, pages 55?66, University of Queensland,Australia, 1988.[26] R. Fonseca, G. Porter, R. H. Katz, S. Shenker, and I. Stoica.X-Trace: A pervasive network tracing framework. In NSDI,2007.[27] D. Geels, G. Altekar, P. Maniatis, T. Roscoe, and I. Stoica.Friday: Global comprehension for distributed replay. In NSDI,2007.[28] C. Ghezzi, M. Pezz?, M. Sama, and G. Tamburrelli. MiningBehavior Models from User-intensive Web Applications. InICSE, 2014.[29] D. Giannakopoulou and J. Magee. Fluent Model Checking forEvent-Based Systems. In FSE, 2003.[30] E. M. Gold. Language Identification in the Limit. Informationand Control, 10(5):447?474, 1967.[31] D. Harel, H. Kugler, and A. Pnueli. Synthesis Revisited: Gen-erating Statechart Models from Scenario-Based Requirements.Form. Meth. in Soft. and Sys. Modeling, 3393, 2005.[32] A. Heussner, T. Gall, and G. Sutre. Extrapolation-Based PathInvariants for Abstraction Refinement of Fifo Systems. In SPIN,2009.[33] A. Heussner, T. Le Gall, and G. Sutre. McScM: A GeneralFramework for the Verification of Communicating Machines.In TACAS, 2012.[34] G. J. Holzmann. The Model Checker SPIN. IEEE Trans. Softw.Eng., 23(5):279?295, May 1997.[35] J. E. Hopcroft, R. Motwani, and J. D. Ullman. Introductionto Automata Theory, Languages, and Computation - (2. ed.).2001.[36] G. Jiang, H. Chen, C. Ungureanu, and K. Yoshihira. Multi-resolution Abnormal Trace Detection Using Varied-length N-grams and Automata. In ICAC, 2005.13University of British Columbia Technical Report[37] I. Krka, Y. Brun, G. Edwards, and N. Medvidovic. Synthesiz-ing Partial Component-Level Behavior Models from SystemSpecifications. In FSE, 2009.[38] S. Kumar, S.-C. Khoo, A. Roychoudhury, and D. Lo. MiningMessage Sequence Graphs. In ICSE, pages 91?100, 2011.[39] S. Kumar, S.-C. Khoo, A. Roychoudhury, and D. Lo. InferringClass Level Specifications for Distributed Systems. In ICSE,2012.[40] L. Lamport. Time, Clocks, and the Ordering of Events in aDistributed System. CACM, 21(7):558?565, July 1978.[41] L. Lin and M. D. Ernst. Improving adaptability via programsteering. In ISSTA, pages 206?216, July 2004.[42] X. Liu, Z. Guo, X. Wang, F. Chen, X. Lian, J. Tang, M. Wu,M. F. Kaashoek, and Z. Zhang. D3S: debugging deployeddistributed systems. In NSDI, 2008.[43] D. Lo, L. Mariani, and M. Pezz?. Automatic Steering of Be-havioral Model Inference. In ESEC/FSE, 2009.[44] D. Lorenzoli, L. Mariani, and M. Pezz?. Automatic generationof software behavioral models. In ICSE, 2008.[45] J. G. Lou, Q. Fu, S. Yang, Y. Xu, and J. Li. Mining Invariantsfrom Console Logs for System Problem Detection. In ATC,2010.[46] F. Mattern. Virtual Time and Global States of Distributed Sys-tems. In Parallel and Distributed Algorithms, pages 215?226,1989.[47] T. Ohmann, K. Thai, I. Beschastnikh, and Y. Brun. MiningPrecise Performance-Aware Behavioral Models from ExistingInstrumentation. In ICSE NIER, 2014.[48] J. H. Perkins, S. Kim, S. Larsen, S. Amarasinghe, J. Bachrach,M. Carbin, C. Pacheco, F. Sherwood, S. Sidiroglou, G. Sul-livan, W.-F. Wong, Y. Zibin, M. D. Ernst, and M. Rinard.Automatically patching errors in deployed software. In SOSP,2009.[49] L. Pitt and M. K. Warmuth. The Minimum Consistent DFAProblem Cannot be Approximated Within any Polynomial. J.ACM, 40(1):95?142, 1993.[50] S. P. Reiss and M. Renieris. Encoding Program Executions. InICSE, 2001.[51] P. Reynolds, C. Killian, J. L. Wiener, J. C. Mogul, M. A. Shah,and A. Vahdat. Pip: detecting the unexpected in distributedsystems. In NSDI, 2006.[52] L. Rizzo. Dummynet: a Simple Approach to the Evaluation ofNetwork Protocols. CCR, 27(1):31?41, Jan. 1997.[53] R. R. Sambasivan, A. X. Zheng, M. D. Rosa, E. Krevat, S. Whit-man, M. Stroucken, W. Wang, L. Xu, and G. R. Ganger. Diag-nosing Performance Changes by Comparing Request Fows. InNSDI, 2011.[54] J. Subhlok and Q. Xu. Automatic construction of coordinatedperformance skeletons. In IPDPS, 2008.[55] A. S. Tanenbaum and D. J. Wetherall. Computer Networks (5thEdition). Prentice Hall, 2010.[56] H. Ural and H. Yenig?n. Towards Design Recovery from Ob-servations, volume 3235, chapter 9, pages 133?149. SpringerBerlin Heidelberg, 2004.[57] Voldemort. http://project-voldemort.com, 2013.[58] W. Xu, L. Huang, A. Fox, D. Patterson, and M. I. Jordan.Detecting Large-Scale System Problems by Mining ConsoleLogs. In SOSP, 2009.[59] M. Yabandeh, N. Knezevic, D. Kostic, and V. Kuncak. Crys-talBall: predicting and preventing inconsistencies in deployeddistributed systems. In NSDI, 2009.[60] J. Yang, D. Evans, D. Bhardwaj, T. Bhat, and M. Das. Perra-cotta: Mining Temporal API Rules from Imperfect Traces. InICSE, 2006.14


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



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"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items