In traditional uniprocessor systems, processor scheduling is the responsibility of the operating system. In high performance computing (HPC) domains that largely involve parallel processors, the responsibility of scheduling is usually left to the applications. So far, parallel computing has been confined to a small group of specialized HPC users. In this context, the hardware, operating system, and the applications have been mostly designed independently with minimal interactions. As the multicore processors are becoming the norm, parallel programming is expected to emerge as the mainstream software development approach. This new trend poses several challenges including performance, power management, system utilization, and predictable response. Such a demand is hard to meet without the cooperation from hardware, operating system, and applications. Particularly, an efficient scheduling of cores to the application threads is fundamentally important in assuring the above mentioned characteristics. We believe, operating system requires to take a larger responsibility in ensuring efficient multicore scheduling of application threads. To study the performance of a new scheduling algorithm for the future multicore systems with hundreds and thousands of cores, we need a flexible scheduling simulation testbed. Designing such a multicore scheduling simulation testbed and illustrating its functionality by studying some well known scheduling algorithms Linux and Solaris are the main contributions of this thesis. In addition to studying Linux and Solaris scheduling algorithms, we demonstrate the power, flexibility, and use of the proposed scheduling testbed by simulating two popular gang scheduling algorithms - adaptive first-come-first-served (AFCFS) and largest gang first served (LGFS). As a result of this performance study, we designed a new gang scheduling algorithm and we compared its performance with AFCFS. The proposed scheduling simulation testbed is developed using Java and expected to be released for public use.
展开▼
