MACT:高通量众核处理器离散访存请求批量处理机制

发布时间：2018-04-04 23:09

本文选题：高通量处理器　切入点：访存请求收集表　出处：《计算机研究与发展》2015年06期

【摘要】：网络服务等新型高通量应用的迅速兴起给传统处理器设计带来了巨大的挑战.高通量众核处理器作为面向此类应用的新型处理器结构成为研究热点.然而,随着片上处理核数量的剧增,加之高通量应用的数据密集型特点,"存储墙"问题进一步加剧.通过分析高通量应用访存行为,发现此类应用存在着大量的细粒度访存,降低了访存带宽的有效利用率.基于此分析,在高通量处理器设计中通过添加访存请求收集表(memory access collection table,MACT)硬件机制,结合消息式内存机制,用于收集离散的访存请求并进行批量处理.MACT硬件机制的实现,提高了访存带宽的有效利用率,同时也提高了执行效率;并通过时间窗口机制,确保访存请求在最晚期限之前发送出去,保证任务的实时性.实验以典型高通量应用WordCount,TeraSort,Search为基准测试程序.添加MACT硬件机制后,访存数量减少约49%,访存带宽提高约24%,平均执行速度提高约89%.
[Abstract]:The rapid rise of new high-throughput applications such as network services brings great challenges to the traditional processor design.As a new processor architecture for such applications, high throughput multicore processor has become a research hotspot.However, with the rapid increase in the number of cores processed on a chip and the data-intensive characteristics of high-throughput applications, the "storage wall" problem is exacerbated.By analyzing the memory access behavior of high-throughput applications, it is found that there are a large number of fine-grained memory access in such applications, which reduces the effective utilization of access bandwidth.Based on this analysis, the memory access collection table MACT-based hardware mechanism is added to the design of high throughput processor, and combined with the message memory mechanism, it is used to collect discrete memory requests and implement the batch processing of .MACT hardware mechanism.The efficiency of accessing bandwidth is improved, and the execution efficiency is improved, and the time window mechanism is used to ensure that the access request is sent out before the latest deadline to ensure the real-time performance of the task.The experiment is based on the typical high throughput application of WordCountl Tera sort search as a benchmark program.After adding the MACT hardware mechanism, the number of access memory is reduced by 49%, the access bandwidth is increased about 24%, and the average execution speed is increased about 89%.
【作者单位】：计算机体系结构国家重点实验室(中国科学院计算技术研究所);中国科学院大学计算机与控制学院;数学工程与先进计算国家重点实验室;中国科学技术大学计算机科学与技术学院;
【基金】：国家“九七三”重点基础研究发展计划基金项目(2011CB302501) 国家“八六三”高技术研究发展计划基金项目(2012AA010901,2015AA011204) “核高基”国家科技重大专项基金项目(2013ZX0102-8001-001-001) 国家自然科学基金项目(61173007,61332009,61204047)
【分类号】：TP332

【参考文献】