Extensible Resource Management for Networked Virtual Computing

Abstract

Advances in server virtualization offer new mechanisms to provideresource management for shared server infrastructures. Resourcesharing requires coordination across self-interested systemparticipants (e.g., providers from different administrative domains orthird-party brokering intermediaries). Assignments of the sharedinfrastructure must be fluid and adaptive to meet the dynamic demandsof clients. This thesis addresses the hypothesis that a new, foundational layerfor virtual computing is sufficiently powerful to support a diversityof resource management needs in a general and uniform manner.Incorporating resource management at a lower virtual computing layerprovides the ability to dynamically share server infrastructurebetween multiple hosted software environments (e.g., grid computingmiddleware and job execution systems). Resource assignments withinthe virtual layer occur through a lease abstraction, and extensiblepolicy modules define management functions. This research makes thefollowing contributions:

Defines the foundation for resource management in a virtual computinglayer. Defines protocols and extensible interfaces for formulatingresource contracts between system participants. Separates resourcemanagement functionalities across infrastructure providers,application controllers, and brokering intermediaries, and explores theimplications and limitations of this structure.

Demonstrates policy extensibility by implementing a virtualcomputing layer prototype, Shirako, and evaluating a range of resource arbitration policies for various objectives. Provides results with proportional share, priority, worst-fit, andmulti-dimensional resource slivering.

Defines a proportional share policy, WINKS, that integrates a fairqueuing algorithm with a calendar scheduler. Provides a comprehensiveset of features and extensions for virtual computing systems (e.g.,requests for multiple resources, advance reservations,multi-dimensional allocation, and dynamic resource pools). Shows thepolicy preserves fairness properties across queue transformations andcalendar operations needed to implement these extensions.

Explores at what layer, and at what granularity, decisions about resource control should occur. Shows that resource management at a lower layer can expose dynamic resource control to hosted middleware,at a modest cost in fidelity to the goals of the policy.