Title of dissertation : EXPLOITING IDLE CYCLES IN NETWORS OF WORKSTATIONS

Title of dissertation: EXPLOITING IDLE CYCLES IN NETWORS OF WORKSTATIONS Kyung Dong Ryu, Doctor of Philosophy, 2001 Dissertation directed by: Associate Professor Jeffrey K. Hollingsworth Department of Computer Science Studies have shown that workstations are idle a significant fraction of the time. Traditional idle resource harvesting systems define a social contract that permits guest jobs to run only when a workstation is idle. To enforce this contract, guest jobs are stopped and migrated as soon as the owner resumes use of their machines. However, such systems waste many opportunities to exploit idle cycles because of overly conservative estimates of resource contention. In this thesis, we present a new policy, called Linger-Longer, that refines the social contract to permit fine-grain cycle stealing. Linger-Longer allows guest jobs to linger on a machine at low priority even when local tasks are active. Also, we developed a new adaptive job migration scheme based on runtime cost/benefit analysis. Our simulation study shows that the Linger-Longer policy can improve the throughput of guest jobs on a cluster by up to 60% with only a few percent slowdown of local jobs. The simulation also demonstrates that guest parallel jobs can perform better with our new approach than with the traditional run-time reconfiguration approach. To limit the impact of guest jobs' resource use, new local resource scheduling policies and mechanisms are required. We present a suite of mechanisms to support prioritized use of CPU, memory, I/O and network bandwidth. An overall prototype of the Linger-Longer system has been implemented in Linux. The prototype integrates the operating system extensions for resource throttling and a new adaptive migration policy module while leveraging general job scheduling and checkpointing mechanisms of an existing system. Using this prototype, we conduct a head-to-head performance comparison between our fine-grain cycle stealing policies and the traditional coarse-grain cycle stealing policies. The experiment on a Linux cluster with a set of benchmark applications show that, overall, Linger-Longer can improve the guest job throughput by 50% to 70%, with only a 3% host job slowdown. EXPLOITING IDLE CYCLES IN NETWORKS OF WORKSTATIONS