LOMARC: Look ahead matchmaking for multi-resource coscheduling.
Date of Award
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.
Hyper-Threading (HT) provides a new possibility for job coscheduling without context switch and without the cost for coordinating processes of one parallel job. However, HT achieves high processor throughput at the expense of reducing the performance of the individual process. Since the hardware resources are actually shared between two coscheduled jobs, the resource contention will harm the performance of each job. Most scheduling approaches only focus on the CPU without considering the impact on other resources. In this thesis we present LOMARC, a space-time sharing approach that takes multiple resources, including CPU, I/O, memory and network, into consideration for job coscheduling on HT processors. To improve resource utilization and reduce job response times, LOMARC matches two jobs with complementary resource requirements to coschedule. Our approach partially reorders the waiting job queue by lookahead to increase the possibility of finding a good match. LOMARC also generalizes for standard CPUs, using an adjusted matching scheme and only focusing on hiding I/O latency. In addition, LOMARC incorporates standard scheduling approaches such as priority ordering, aging and backfilling. In our simulation experiment, we use a realistic workload model to provide the convincing results. Our experimental results demonstrate that LOMARC delivers better performance than the standard space sharing approach and the other two job coscheduling approaches for HT processors. The performance gain is mainly due to an increased possibility of coscheduling two complementary jobs by looking ahead on the waiting queue. Source: Masters Abstracts International, Volume: 43-01, page: 0239. Adviser: Angela Sodan. Thesis (M.Sc.)--University of Windsor (Canada), 2004.
Lan, Lei., "LOMARC: Look ahead matchmaking for multi-resource coscheduling." (2004). Electronic Theses and Dissertations. 2641.