Java虚拟机在Intel众核架构下的动态数据预取研究与优化

摘要

Intel Xeon Phi coprocessor,as one of the most representative many-core products,provides very powerful hardware support and computing resources.However,the memory design that Xeon Phi employs has rather high access latency,thus relies heavily on cache data prefetching techniques for performance improvement.Java,as a widely-used language with automatic memory management,currently does not have such memory-related optimizations for Xeon Phi architecture.In this paper,we perform a comprehensive study on Xeon Phi's prefetching mechanism and propose a dynamic data prefetching solution with an implementation inside HotSpot virtual machine.Compared to traditional static methods and current dynamic prefetching schemes,our solution is more suitable for Xeon Phi many-core architecture and Java dynamic runtime environment.The evaluation results demonstrate that our solution could achieve an averagely 2.5x performance speedup and improve the best multi-threaded throughput by 40％ on Xeon Phi.%Intel Xeon Phi协处理器作为现阶段极具代表性的众核产品之一,为应用程序提供了强大的硬件环境和计算资源.然而,Xeon Phi所采用的内存设计存在高访问延迟的问题,因此非常依赖于缓存数据预取技术以提升访存性能.而Java作为一门具有自动内存管理且被广泛使用的语言,现有设计并未针对于Xeon Phi架构采取访存相关的优化.本文详细地研究了XeonPhi上的缓存预取机制,并在HotSpot虚拟机内部设计实现了一套动态的运行时缓存预取解决方案,该方案相比传统的静态方法和现有动态预取方案更适合于Xeon Phi众核架构及Java动态语言环境.本文通过实验表明,该动态预取方案在Xeon Phi众核平台上可以带来平均2.5倍的单线程加速比以及40％的多线程最优性能提升.