摘要
异构计算是高效能计算发展的必然趋势,针对异构计算运行中并行任务和体系结构难匹配的问题,提出了实现并行任务和体系结构匹配的并行任务分簇方法。首先给出效能的概念及异构计算中体系结构感知的分簇问题,然后从理论上分析了异构匹配与效能的关系,提出了实现异构计算匹配和结构匹配的分簇理论,目的是发挥异构计算中机器的潜能,协同处理并行任务,实现高效能。在此基础上,给出相应的算法。最后通过仿真实验说明,该方法可通过簇图与体系结构的匹配缩短通信开销在执行时间上所占的比例,从而缩短并行执行时间,以提高系统利用率,最终实现异构计算的高效能。
Heterogeneous computing has been a trend of high-productivity computing. Matching between parallel task and architecture in heterogeneous computing becomes a key idea to realize high productivity. We provided parallel task clustering policy based on matching between parallel task and architecture. Firstly we gave the concept of high-produe- tivity and the problem of clustering on heterogeneous computing. Secondly after theoretically analyzing the relation be- tween heterogeneous matching and produetivity, we gave the method of realizing respectively computing and structure matching. Thirdly we gave accordingly the architecture-aware parallel task clustering algorithm. Finally the simulation experimental results show that such algorithms can effectively realize heterogeneous matching and enhance the hetero- geneous computing productivity.
出处
《计算机科学》
CSCD
北大核心
2013年第3期121-125,共5页
Computer Science
基金
国家863高技术研究发展计划(61103068)
NSFC-微软亚洲研究院联合项目(2009AA012201)
国家自然基金项目(60970155)
教育部博士点基金项目(20090072110035)
上海市优秀学科带头人计划项目(10XD1404400)
高效能服务器和存储技术国家重点实验室开放基金项目(2009HSSA06)资助
关键词
异构计算
并行任务匹配
体系结构感知
分簇
Heterogeneous computing, Parallel task matching, Architecture-aware, Clustering
