一种多线程程序内存系统模拟器Trace驱动仿真方法

发布时间：2018-12-17 04:43

【摘要】：伴随大数据计算时代的到来,片上多核处理器为提高多线程程序服务器吞吐率发挥巨大作用,同时其内存系统的访问延迟越来越影响系统性能.目前,路径驱动(trace-driven)仿真方法比执行驱动(execution-driven)运行速度快,被内存系统研究者广泛采用.但是路径驱动在仿真并发线程时,会同时导致宏观和微观的访存错位.而实际多线程程序运行过程中,不会发生这种访存错位行为.通过理论分析和计算,访存错位引起路径驱动的仿真结果存在明显偏差.针对上述问题,提出了一种方法来避免路径驱动仿真发生宏观和微观访存错位,精确回放采集阶段的多线程程序行为.实验数据显示,在避免宏观访存trace错位后,多线程程序的多个仿真指标出现最高10.22%的变化;对于部分访存密集型的多线程程序,避免微观访存trace错位可以使算数平均IPC出现大于50%的变化.为研究交互线程的内存系统行为提供一种更加准确的路径驱动方法.
[Abstract]:With the advent of big data computing era, the multi-core processor on chip plays a great role in improving the throughput of multithreaded program server, and the memory system access delay is increasingly affecting the performance of the system. At present, the path driver (trace-driven) simulation method is faster than the executive driver (execution-driven), and is widely used by memory system researchers. However, path drivers can lead to both macro and micro memory access mislocations when simulating concurrent threads. And the actual multi-thread program running process, will not occur this memory-access dislocation behavior. Through theoretical analysis and calculation, there is a significant deviation in the path driven simulation results caused by memory access misalignment. To solve the above problems, a method is proposed to avoid macro and micro memory access mislocations in path-driven simulation, and to accurately play back the multi-thread program behavior in the acquisition stage. The experimental data show that after avoiding the malposition of trace, the multithreaded program has the highest change of 10.22%. For some memory-intensive multithreading programs, avoiding the misalignment of microcosmic memory access trace can make the average arithmetic IPC change by more than 50%. It provides a more accurate path driving method for studying the memory system behavior of interactive threads.
【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;中国科学院计算技术研究所先进计算机系统研究中心;
【基金】：国家自然科学基金项目(61272132,61221062)
【分类号】：TP333.1

【参考文献】