众核处理器片上同步机制和评估方法研究  

On Synchronization and Evaluation Method of Chipped Many-Core Processor

作　　者：徐卫志[1,2] 宋风龙[1] 刘志勇[1] 范东睿[1] 余磊[1,2] 张帅[1,2] 

机构地区：[1]中国科学院计算技术研究所系统结构重点实验室,北京100190 [2]中国科学院研究生院,北京100039

出　　处：《计算机学报》2010年第10期1777-1787,共11页Chinese Journal of Computers

基　　金：国家自然科学基金重点项目(60736012);国家"九七三"重点基础研究发展规划项目基金(2005CB321600);国家"八六三"高技术研究发展计划项目基金(2009AA01Z103);国家杰出青年科学基金(60925009);国家自然科学基金创新研究群体科学基金(60921002);北京市自然科学基金(4092044)资助~~

摘　　要：同步机制是片上多核/众核处理器正确执行和协同通信的关键,其效率对处理器的性能非常重要.针对片上众核体系结构,提出并实现了两种粗粒度同步机制和一种细粒度同步机制,即片上专用硬件支持的同步机制、基于原语的片上互斥访问同步机制和基于满空标志位的细粒度同步机制;提出了粗粒度同步机制的评估标准和评估方法,并设计了量化评估程序.以片上同构众核处理器Godson-T模拟器和AMDOpteron商业片上多核处理器为平台,评估比较了提出的硬件支持的同步机制与基于原语的同步机制的性能.结果表明,硬件支持可以使得片上众核处理器的同步机制性能明显提高;在传统基于原语的同步机制中,大部分性能损失是由于负载不平衡和同步点的串行化操作而造成的等待时间.Synchronization schemes are critical for execute correctly and communicate cooperatively. on-chip multi-core and many core processor to The efficiency of the synchronization is very important for the processor. In this paper, for on-chip many-core architecture, three types of synchronization schemes are proposed. That is, two types o[ coarse-grain synchronization schemes based on dedicated hardware support and atomic operation, and a fine-grain synchroniza- tion scheme based on Full/Empty bit. Then, the evaluation criterions and methods are proposed, in which quantitative evaluation micro benchmarks are designed for coarse-grain synchronization schemes. Finally, the coarse-grain synchronization schemes are evaluated via a many-core archi- tecture simulator, i. e. , Godson-T, and AMD Opteron commercial on-chip multi-processor using pThread multi-thread program model. The results show that hardware support improves the per- formance of the synchronization obviously for on-chip many-core processor, and the performance loss of the traditional synchronization scheme based on atomic instructions is caused by the wait- ing cost of load imbalance and serialization on synchronization point mostly.

关 键 词：片上众核处理器 同步 硬件支持 量化评估 微程序 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]