On the Design and Implementation of an Effective Prefetch Strategy for DSM Systems

  • Authors:

  • Hsiao-Hsi Wang;Kuan-Ching Li;Kuo-Jen Wang;Ssu-Hsuan Lu

  • Affiliations:

  • Parallel and Distributed Processing Center, Dept. of Computer Science and Information Management, Providence University, Taichung, Taiwan 43301;Parallel and Distributed Processing Center, Dept. of Computer Science and Information Management, Providence University, Taichung, Taiwan 43301;Parallel and Distributed Processing Center, Dept. of Computer Science and Information Management, Providence University, Taichung, Taiwan 43301;Parallel and Distributed Processing Center, Dept. of Computer Science and Information Management, Providence University, Taichung, Taiwan 43301

  • Venue:
  • The Journal of Supercomputing
  • Year:
  • 2006

Quantified Score

Hi-index 0.00

Visualization

Abstract

Distributed Shared Memory systems provide the abstraction of a shared address space among computing hosts interconnected via a private network, in a convenient and easy way to achieve high performance. However, there are several drawbacks to these systems, such as communication overhead, network latency, false sharing, coherence and page faults. By prefetching, one can overlap communication and computation though the Accumulated Waiting Phenomenon and the Waiting Synchronization Phenomenon affect overall performance. In this paper, an Effective Prefetch Strategy (EPS) is proposed, to improve the shortcomings of previous prefetching approaches and increases the prefetch page hit rate. In addition, our EPS strategy reduces the waiting time for each computing host upon barrier synchronizations and misprefetches. Experimental comparisons show that our proposed EPS strategy offers the best performance among existinrefetching strategies.