Loop-Level Process Control: An effective processor allocation policy for multiprogrammed shared-memory multiprocessors

Abstract

Processor time-sharing is the most common way to increase the overall system utilization for shared-memory multiprocessor systems. However, the performance of individual applications might be sacrificed due to the high overhead of context switching, due to the processing power wasted by busy-waiting synchronization and locking operations, and due to poor cache memory utilization. In this paper, we propose a simple and effective processor allocation scheme, called Loop-Level Process Control (LLPC), for multiprogrammed multiprocessors. At the beginning of each parallel section of each application program, LLPC uses the current system load to determine an upper limit on the number of processes the application can create for that parallel section. Preliminary simulation results using the Perfect Club Fortran benchmarks show that this loop-level process control scheme can produce a high system utilization while maintaining high performance for the individual applications. Another advantage of this strategy is that it is transparent to the programmer and does not require any modifications to the operating system. Consequently, the application can remain portable and compatible.