Mala Ghanesh
Department of Computer Science
University of Houston
e-mail : mghanesh@cs.uh.edu
Research Advisor : Dr. Jaspal Subhlok
Abstract of Research :
Parallel systems are normally space shared, or employ gang scheduling to run multiple jobs on the same set of nodes. Independent scheduling of application threads on system nodes offers more flexibility but the overhead is generally considered unacceptable. In this research, we experimentally show that the performance under independent scheduling is good and competitive with space sharing and gang scheduling for a large class of applications on small clusters. The key reason is that CPU scheduling policies implicitly minimize the overhead of asynchronous execution. Independent scheduling can provide acceptable application performance with sharing, along with other benefits like flexibility and better resource utilization, when compared to gang scheduling and space sharing. Under independent scheduling, performance of a parallel application with sharing depends on application characteristics and operating system features, beside the sharing scenario. Significant application factors are CPU utilization and the communication volume and frequency. The node scheduling policy and CPU time slice are significant operating system factors. All experiments are done with NAS benchmarks on a small Linux cluster.