返回Google圖書搜尋

Techniques for Running Synchronous Parallel Programs on a Computational Grid

出版	University of North Carolina at Charlotte, 2007
URL	http://books.google.com.hk/books?id=krJlnQEACAAJ&hl=&source=gbs_api

註釋Grid computing allows multiple geographically distributed organizations to share resources across the Internet. This technology benefits diverse sectors such as government, science and industry. However, it also poses problems to porting some types of parallel programs that would suffer from some of the Grid's disadvantages such as high latencies and a heterogeneous environment. This thesis studies the feasibility to port embarrassingly parallel, pipeline, and synchronous parallel computing techniques to the Grid. It reviews latency hiding to improve synchronous parallel programs and introduces and tests the use of a technique called latency hiding with redundant processing (LHRP) to exchange processor power and memory for a lower perceived WAN latency. LHRP performed better than using latency hiding for jobs where internal compute time is not enough to hide the latency. The thesis also introduces Data Matrix Row-height CPU Performance Adaptation (DMRCPA) and Multi-Border CPU Performance Adaptation MBCP A. Both are load balancing algorithms to improve a program's performance in the Grid environment. DMRCPA was implemented and tested. It performs better than partitioning the data matrix equally among all the processors. The thesis also introduces the use of a model that accurately describes the behavior of lightly-coupled synchronous parallel programs.