摘要
研究了在基于片上网络(Network on Chip,NoC)结构的单芯片多处理器(Chip Multiple Processors,CMPs)中,访存请求的NoC延时差异对存储系统的公平性带来的影响.针对该问题进行了理论分析、抽象,并构建试验模型,从网络规模、报文比例等4个方面对造成访存请求的NoC延时差异的原因进行了讨论.最后提出了一种基于片上网络延时的存储器访问调度方法(Scheduling Based on NoC Latency,SBNL),与传统的方法相比,能够将NoC延时差异对访存请求公平性的影响降低20%左右,并带来15.7%的执行效率提升.
In the CMPs (Chip Multiple Processors) based on NoC (Network on Chip), the NoC latency difference of memory requests would have great effect on the fairness of memory systems. The theoretical and experimental models were constructed and the causations of NoC latency difference were discussed from the scale of network, proportion of pockets and other two aspects. The SBNL (Scheduling Based on NoC Latency) method was proposed, which could reduce the NoC latency difference's side effect on the memory fairness by about 20% and bring 15.7% increment of the execution efficiency, compared with the traditional methods.
出处
《计算机学报》
EI
CSCD
北大核心
2011年第8期1500-1508,共9页
Chinese Journal of Computers
基金
国家"八六三"高技术研究发展计划重大项目(2007AA01Z108)
"核高基"重大专项(2009ZX01034-001-001-006)
国家"八六三"高技术研究发展计划(2009AA011704)
国家自然科学基金(61070036)资助