SimHPC:一种基于执行驱动的大规模并行系统模拟器  

SimHPC:An Execution-Driven Simulator for High-Performance Computers

作　　者：刘轶[1] 支予哲[1] 张昕[2] 李鹤[2] 焦林[2] 张鹏[2] 苏阳明[1] 倪泽辉[1] 钱德沛[1] 

机构地区：[1]北京航空航天大学中德联合软件研究所,北京100191 [2]西安交通大学计算机系,西安710049

出　　处：《计算机学报》2013年第4期738-746,共9页Chinese Journal of Computers

基　　金：国家自然科学基金(61073011;61133004);国家"八六三"高技术研究发展计划项目基金(2012AA01A302)资助~~

摘　　要：模拟实验方法对高性能计算机系统的性能评价和优化设计有着重要的意义,然而由于目标系统规模庞大,传统的体系结构模拟器难以满足模拟性能方面的要求.文中提出了一种专门用于高性能计算系统的模拟器——SimHPC,该模拟器采用执行驱动的全系统模拟方法,支持操作系统和应用程序的模拟运行.通过采用与目标系统同构的节点作为宿主节点以及并行模拟的方法,使得模拟性能相比传统的体系结构模拟器大幅提高,与现有的几种大规模并行系统模拟器相比,SimHPC在通用性和模拟性能方面也具有一定的优势.Simulation plays an important role in performance evaluation and design improvement of high-performance computing systems.However,due to the huge scale of this kind of machines,traditional architecture simulators such as Simics and M5 cannot meet requirements of simulation efficiency.This paper presents the SimHPC simulator dedicated for high-performance computer systems.As an execution-driven full-system simulator,SimHPC supports startup of Linux system and execution of parallel applications including Linpack.To achieve performance improvement significantly,two methods are used in SimHPC.Firstly,the same nodes with the target machine are used in the host system,this makes the execution time of any instruction sequence in the host node equals to the target,and many time-consuming works can be avoided,such as modeling and simulation of processor microarchitecture and instruction set;secondly,parallel host system are used to simulate target systems,and node number of the host can be adjusted according to the size of target machine and requirements in simulation performance.Experiment results show that the slowdown factor of the simulator is limited to 127 when simulating 1024-nodes parallel system on 4-nodes host system.Compared to existing parallel system simulators,SimHPC also behaves advantages in target system generality and simulation performance.

关 键 词：模拟器 高性能计算 并行系统 执行驱动 性能评价 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]