面向应用的流存储系统评测与改进  

Application-specific Evaluation and Optimization of the Stream Memory System

作　　者：汪芳[1] 安虹[1,2] 徐光[1] 许牧[1] 姚平[1] 

机构地区：[1]中国科学技术大学计算机科学技术系,安徽合肥230027 [2]中国科学院计算机系统结构重点实验室,北京100080

出　　处：《小型微型计算机系统》2010年第5期990-995,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金重点项目(60633040)项目;国家自然科学基金项目(60736012)资助;国家"九七三"重点基础研究发展计划项目(2005CB321601)资助;国家"八六三"高技术研究发展计划重大项目(2006AA01A102-5-2)资助;国家"八六三"高技术研究发展计划项目(2009AA01Z106)资助;国家科技重大专项项目(2009ZX01036-001-002)资助;教育部-英特尔信息技术专项科研基金项目(MOE-INTEL-08-07)资助

摘　　要：有限的片外存储带宽是制约流处理器性能提升的瓶颈之一,流存储系统已经采用了多种方式来缓解这个问题,但当前的设计并没有充分考虑应用具体的访存模式对有效带宽利用率的影响.通过分析和实验,评估流存储系统主要设计参数对不同访存模式的优化效果;在此基础上针对不同的流访问并行度提出了相应的结构改进,加入宽发射和短作业优先调度支持,充分挖掘存储访问的局部性和并行性,改善了负载平衡,从而有效地提高了片外带宽的使用效率和流程序的整体性能.The limited off-chip bandwidth of memory accesses increasingly becomes the bottleneck of entire stream processing system. Many methods have been adopted into stream memory system to alleviate this problem,but current design didn＇t consider enough about the relationship between application-specific memory accessing patterns and the utilization rate of off-chip bandwidth. This paper first estimates the effect of primary design parameters targeted on different access patterns through analysis and experiments. Based on these results,some architecture modifications are proposed for various parallel degrees of stream accesses. By widening the address generators and adding short-task priority scheduling,the locality and parallelism among memory accesses are explored fully,along with better load balance. These optimizations can significantly improve the utilization efficiency of DRAM bandwidth and further boost the final performance of the entire streaming program.

关 键 词：流存储系统 DRAM 存储带宽 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]