一种基于元数据的采样模拟技术优化  

A Metadata-Driven Optimization for Sampling Simulation

作　　者：严强[1] 张为华 刘力力[1] 臧斌宇[1] 朱传琪[1] 

机构地区：[1]复旦大学并行处理研究所,上海201203

出　　处：《计算机学报》2008年第11期1986-1994,共9页Chinese Journal of Computers

基　　金：Intel大学合作计划资助.

摘　　要：分析了目前主流采样模拟技术中定长样本的不足,提出了一种基于编译器元数据的采样模拟技术(BigLoopSP).首先利用编译器收集各种可能的周期行为的边界信息作为元数据.然后为了处理程序中大量存在的动态行为,基于编译器产生的元数据结合程序的动态行为进行周期行为的划分和采样点的选取.以此方案划分的变长候选样本能够在保证样本质量的前提下有效地减少所需特征样本的总数.因此比较于定长采样技术SimPoint,BigLoopSP在提高精确性的同时,进一步降低了模拟所需的时间(相对于SimPoint的平均加速比为2.63).This paper analyzes the disadvantage of mainstream sampling simulation techniques using fixed-length samples and proposes a metadata-driven optimization for sampling simulation,BigLoopSP.In the approach,the compiler selects candidate loops and annotates the boundaries of those loops as metadata.Those metadata are used to divide the execution into varied-length candidate samples,for which each candidate sample corresponds to one iteration of the chosen loop.Since the program execution exhibits dynamic behaviors,the approach combines the knowledge from the metadata and the dynamic profiles to guide phase partition and selects simulation points for those phases.This approach effectively reduces the number of representative samples while preserving the good quality of them.So,compared with those mainstream sampling simulation techniques,such as SimPoint,our approach achieves better accuracy and reduces more simulation time（a speedup of 2.63X over SimPoint）.

关 键 词：编译 模拟器 元数据 采样模拟 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]