Involving Memory Resource Consideration in Workload Distribution for Software DSM Systems

Abstract

This paper is aimed at resolving the problem of workload distribution for software distributed shared memory (DSM) systems. According to the past methods, DSM systems usually took only CPU resource into account for workload distribution. They distributed program threads based on the computational capability of processors to make load balance, expecting thereby the minimal execution time of applications. However, the cost of memory accesses is an important factor for program performance in addition to computational cost. If processors cannot afford enough physical memory space to cache all the data needed by threads, these processors must perform page replacements to make memory space for data caching while executing the threads. The execution of the threads absolutely will be delayed due to the latency of performing page replacement, and then the execution time of programs will increase. Therefore, only accomplishing load balance cannot exactly guarantee the minimal execution time of programs. In this paper, we propose a new workload distributing method, which simultaneously considers CPU resource and memory resource, for DSM systems. Our experimental results show that memory resource is indeed an important consideration for workload distribution on DSM systems, and our method is more effective for minimizing the execution time of DSM applications than the others considering only either CPU resource or memory resource.