基于OpenCL的流式应用程序在MPSoC上的动态并行度伸缩调度  

An openCL based streaming applications program's dynamic parallelism scaling scheduling on MPSoC

作　　者：黄姗[1,2,3] 石晶林[1,2,3] 萧放 

机构地区：[1]中国科学院计算技术研究所无线通信技术研究中心,北京100190 [2]北京市移动计算与新型终端重点实验室,北京100190 [3]中国科学院大学,北京100049

出　　处：《高技术通讯》2016年第12期925-934,共10页Chinese High Technology Letters

基　　金：国家自然科学基金(61431001);北京市青年拔尖人才(2015000021223ZK31)资助项目

摘　　要：分析了嵌入式系统应用程序的复杂化和多样化趋势,面向嵌入式系统常见的流式应用程序,提出了基于开放运算语言(OpenCL)的统一编程框架,并在此框架的基础上设计一个运行时系统,在应用程序可用计算资源发生变化的场景下,该系统可在线调整应用程序的并行度,并进行动态调度。实验结果显示,与已有的Flextream动态调度系统相比,该调度系统在性能上最高可以提场17%,在动态调度的时间开销上最多可以降低7%。The complex and diversity trends of the application programs for embedded computing systems were analyzed. Then, a unified programming framework based on the open computing language （OpenCL） was proposed for em- bedded computing systems＇ common streaming application programs, and on the basis of the framework, a runtime system was designed. Under the variation of application programs＇ computing resources, the system on-line regu- lates programs＇ parallelism, and conducts dynamic parallelism scaling scheduling. The experimental results showed that, compared with the existing dynamic scheduling system of Flextream, the proposed scheduling system＇ s per- formance was improved by 17% , and the runtime overhead of the dynamic scheduling was reduced by 7%.

关 键 词：多处理器片上系统(MPSoC) 开放运算语言(OpenCL) 编程框架 并行度伸缩 运行时系统 

分 类 号：TP368.1[自动化与计算机技术—计算机系统结构]