Open Collections

Iqbal, Syed Mubashir

University of British Columbia

Datacenters have become a ubiquitous solution for users with large compute and storage needs. Most datacenter operators offer two different products, commonly referred to as virtual machines and spot instances, for accessing compute resources on their datacenter servers. Spot instances are a lower-priced alternative to virtual machines that are opportunistically scheduled on servers to utilize any residual capacity; they are evicted when virtual machine demand spikes. There is no prior work that explores whether the virtual machine and spot instance scheduler designs can be modified in a way to achieve greater revenue from spot instances without degrading scheduling quality for virtual machines. This thesis makes the following contributions: (1) A framework for cooperative scheduling of virtual machines and spot instances, in which the schedulers are able to communicate during the request allocation process. The framework allows for easy plug-and-play integration of different allocation and eviction policies into the schedulers. (2) A realistic virtual machine and spot instance request workload. This workload was generated from a production trace released by Microsoft Azure. (3) An evaluation that quantifies the effect on spot instance revenue of schedulers with different allocation/eviction policies and different design parameters. This evaluation shows it is possible to achieve up to a 160.4% improvement, with an average improvement of 43.1% across all experiments, in spot instance revenue with simple changes to scheduler design. These revenue gains are achieved without any virtual machine scheduling failures.

Text; Other

2022-07-31

Attribution 4.0 International

http://hdl.handle.net/2429/79161

Computer Science

University of British Columbia

UBCV

http://creativecommons.org/licenses/by/4.0/