摘要
并行虚拟机(PVM)是当前流行的大型科学计算并行消息传递环境之一,其系统设计思想更多考虑了异构集群下的兼容性,默认性能设置保守。矩阵乘法是并行计算实现效率提升的典型运算之一,对PVM环境下预设并行矩阵乘法实际测试,给出从任务粒度划分、消息封包寻径及任务分派方面提升PVM并行效率的数据。对数据分析可得:PVM任务划分粒度要大,任务粒度以匹配节点数为最优;合理选用消息封包、寻径策略可改善通信能力的瓶颈;增加各节点的有效运算时间是提高整体并行效率的关键。
PVM is one of the most popular parallel programming environments.But the default setting of PVM system is not the best in some specific cases.Matrix multiplication is one of the most typical operations in the fields of computing parallelization.By testing the example of matrix multiplication on PVM platform,this paper provides the test data of PVM performance improvement under different task division,message package/route and task distribution conditions.It is suggested that PVM task granularity should match the scale of cluster.Choosing message package/route strategy properly can improve communication performance.Increasing the effective running time of each node is the key to improve the performance of PVM.
出处
《桂林工学院学报》
北大核心
2010年第1期168-173,共6页
Journal of Guilin University of Technology
基金
广西自然科学基金资助项目(桂科自0832249)
关键词
PVM
任务划分
消息传递
任务分派
集群
性能优化
parallel virtual machine（PVM）
task division
message passing
task distribution
cluster
performance optimum
