项目名称: 多核环境下程序存储局部性检测与预测方法
项目编号: No.61472008
项目类型: 面上项目
立项/批准年度: 2015
项目学科: 自动化技术、计算机技术
项目作者: 罗英伟
作者单位: 北京大学
项目金额: 84万元
中文摘要: 随着多核、多处理器应用的普及,并发程序对共享资源的竞争成为程序运行的瓶颈。由于计算机存储架构变得越加复杂,分布在一台机器上的CPU通常会共享最底层的缓存资源,并共同通过总线对内存及磁盘进行读写操作,从而共用总线带宽。通过分析程序行为特征和程序的周期性,我们可以合理的分配程序对共享资源的占用,通过调整程序对缓存资源、内存资源以及带宽的使用,从而提升并发程序的性能。我们将寻找一个开销低并支持即时检测的测量方法,利用x86架构中的性能检测单元,以及操作系统的页表机制,在线监测程序的行为特征,如程序的存储局部性和周期性,来预测其对缓存、内存和带宽的占用,进而进行合理的调度。这种方法同时也可以被应用到虚拟化系统中,用于调度虚拟机以及对虚拟机的迁移做出决策。
中文关键词: 多核;程序局部性;测量;预测
英文摘要: As multi-core and multi-processor become commonplace, the competition of shared resource between concurrent programs has become the bottleneck of program performance. Due to the complexity of computer architecture, the CPUs distribute on one machine will share the last level cache (LLC) and read/write memory and disk through the same bus. By analyzing the behavior and periodicity of a program, we could intelligently assign shared resource to the concurrent programs, and schedule their usage of shared cache, memory and bandwidth, so as to improve their performance. We'll seek for a measurement with online-availability and low cost, by the benefit of Performance Monitor Unit (PMU) in x86 architecture, and also the mechanism of page tables, to use it to direct the scheduling via predicting one program's locality and periodicity. This measurement and scheduling policy will be applied onto virtualized environment to manage the virtual machines and their migration.
英文关键词: Multicore;Program Locality;Measure;Predict