多核环境下编译器辅助消息驱动的动态调度  

Compiler-Assisted Message-Driven Dynamic Scheduling on Multicore Processor

作　　者：于俊清[1,2] 余华飞[1] 魏海涛[1] 秦明康[1] 

机构地区：[1]华中科技大学计算机科学与技术学院,武汉430074 [2]华中科技大学网络与计算中心,武汉430074

出　　处：《计算机学报》2014年第7期1633-1637,共5页Chinese Journal of Computers

基　　金：国家"八六三"高技术研究发展计划重点项目基金(2012AA010902);高等学校博士学科点专项科研基金(20120142110089);中国科学院计算技术研究所国家重点实验室开放基金(ICT-ARCH200804)资助~~

摘　　要：计算密集型应用中存在大量的并行性,如何优化并行处理任务,充分利用多核处理器的并行资源是当前多核编程领域亟待解决的重要问题.针对此问题,文中提出编译器辅助消息驱动的动态调度模型来优化任务调度;设计出新的编程模型DFBrook,以便更好地支持数据级和任务级的并行;采用编译器辅助的方式,由编译器产生将在运行时创建数据流图的代码;调度器对数据流图进行分析和解除任务依赖,并将任务分发到各处理核,处理核执行完任务后提供反馈信息给主控核进行调度决策.实验结果表明,调度模型具有较好的扩展性,并保持了较高的负载均衡效率和较低的通信开销.There are abundant parallelisms in compute- intensive applications. Currently, the most important issue is how to optimally process tasks in parallel and make the best use of multi- core processor＇s parallel resources in multi-core programming. In order to tack with this problem, a compiler-assisted message-driven dynamic scheduling model is proposed to orchestrate task scheduling. DFBrook is designed in order to supports data-level and task-level parallelism in a better way. And a compiler-assisted method is adopted to generate data flow graph construction code which will be executed in runtime to create data flow graph. The scheduler analyzes the data flow graph to get rid of dependences between tasks, and then distributes tasks to processing cores. Processing cores pass the feeding back information to the control core for scheduling determination efficiently after finishing tasks. Experimental results demonstrate that the scheduling model is scalable with processor cores while keeping high load balance efficiency and low communication overhead.

关 键 词：多核处理器 单程序多数据 动态调度 消息驱动 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]