结合PVT模拟和排队模型的系统级主存性能分析

Combining PVT Simulation and Queuing Model for System-Level Performance Analysis of Main Memory System

由于主存系统的性能受到多种因素的影响,现有方法不能快速地得到可靠的分析结果,从而影响芯片质量和上市时机.为解决此问题,提出将带时序的程序员视图(PVT)模拟和排队论相结合的方法——ComPQ.首先从PVT模拟中提取与访存相关的系统级实时参数,然后将主存系统抽象为非抢占优先的M/G/1排队模型,再结合实时参数进行性能分析,得到平均访存延迟结果.由于PVT的建模和模拟代价小,从中得到的实时参数弥补了静态理论分析的不足;同时,排队论也提高了纯PVT模拟的精度.实验结果表明,ComPQ与周期精确级模拟相比平均误差为6.38%,最后用主存系统设计空间探索的实例验证了ComPQ的有效性.

Main memory system performance is sensitive to many parameters.Existing analysis approaches are time-consuming and unreliable that can not eliminate performance bottlenecks under time-to-market pressures.To solve this problem,this paper presents a hybrid analysis approach called ComPQ,which combines programmer＇s view with timing（PVT） simulation with queuing model.ComPQ first collects system-level real-time parameters from PVT simulation.Then an MG1 queuing model with non-preemptive priority is combined with the parameters to determine the average memory access latency.The fast PVT simulation and the queuing model increase the accuracy of each other.Experimental results show that compared with cycle accurate model,the average performance error of ComPQ is 6.38%.Finally,we present main memory system space exploration case studies to demonstrate the effectiveness of using ComPQ for exploration.