项目名称: 多核平台上的动态访存优化
项目编号: No.61303051
项目类型: 青年科学基金项目
立项/批准年度: 2014
项目学科: 自动化技术、计算机技术
项目作者: 王振江
作者单位: 中国科学院计算技术研究所
项目金额: 27万元
中文摘要: 长期以来,处理器和内存间的速度差异一直是计算机系统的主要性能瓶颈。多核平台增强了总运算能力却没有同比提高访存带宽,使带宽又成为另一个瓶颈。分析程序在运行时的访存行为,利用动态优化技术缓解访存"慢"和"挤"的问题,是具有重要研究价值的课题。 项目准备在以下三点寻求突破:(1)多线程内存池优化。通过动态运行信息抓取和数据布局,在改善数据局部性的同时,适应多线程程序的不同访存模式,检测、预测和处理假共享。(2)数据动态重布局。研究运行时安全、低开销的移动数据的方法,适应不断变化的热访存序列。(3)访存竞争模型和任务调度算法。为弥补单一数据指标的偏差和局限性,研究综合了多项指标的访存竞争模型,为任务调度等优化提供更好的指导。 研究成果将从分配时和分配后的数据布局、访存竞争等不同角度缓解存储墙问题,提高计算机系统的整体性能。
中文关键词: 多核;访存优化;动态优化;剖析;调试
英文摘要: The speed gap between process and memory has long been the main performance bottleneck of computer system. Multi-core architecture improves the process capability, but not the memory bandwidth, making the latter another bottleneck. Analyzing runtime memory behavior and relieving the memory problems by dynamic optimization is of great research value. We will research on the following issues. (1) Multi-threaded memory pool allocation. It profiles runtime information and manages data layout at allocation time. Besides, it can fit various memory access characteristics in multi-threaded programs, and can detect, predict and handle false sharing. (2) Dynamic data layout rearrangement. It can safely move data at runtime with little overhead when hot memory access sequence changes. (3) Memory contention model and scheduling algorithm. Instead of a limited single metric, the model uses several metrics and describes the degree of memory contention. It can provide better direction to task scheduling as well as some other optimizations. The research targets the memory wall problem from various aspects, including allocation time, post allocation time, and memory contention. The research will improve the overall performance of computer system.
英文关键词: multi-core;memory optimization;dynamic optimization;profiling;debugging