软硬件协同的远端内存系统研究  

Hardware-software co-designed remote memory system

作　　者：李海锋 刘珂[1] 陈明宇[1,2] LI Haifeng;LIU Ke;CHEN Mingyu(Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190;University of Chinese Academy of Sciences,Beijing 100049)

机构地区：[1]中国科学院计算技术研究所,北京100190 [2]中国科学院大学,北京100049

出　　处：《高技术通讯》2023年第10期1009-1019,共11页Chinese High Technology Letters

基　　金：国家重点研发计划(2022YFB4500401);国家自然科学基金(62072439);北京市自然科学基金(212028);山东省自然科学基金联合基金(ZR2019LZH004)资助项目。

摘　　要：针对当前远端内存系统中页面预取与页面替换因操作系统与应用程序之间语义鸿沟导致的局限性问题,本文提出一个软硬件协同的远端内存系统。通过在内存控制器中增加热点页面提取表,将实时访存的热点页面信息通过内存中的缓冲区传送给操作系统。同时,通过对访存信息的学习,构建了高精度的异步预取框架与替换框架,降低应用关键数据路径的开销,提升远端内存系统的性能。本文利用内存跟踪工具构建了一个原型仿真系统。实验证明,在拥有全局实时访存信息后,预取框架可以实现超过90%的准确率与覆盖率,与谷歌提出的远端内存系统Fastswap相比,性能提升59%。相比于内核默认替换框架,替换框架使应用性能提升30%。Currently,page prefetching and page replacement in remote memory systems are limited due to the semantic gap between the operating system and the application access behavior.This paper proposes a hardware and software co-designed remote memory system to bridge the gap.A hot page detection in the memory controller is added to transfer real-time memory access information to the operation system.At the same time,through the learning of memory access information,an asynchronous prefetching framework and replacement framework is built to reduce the cost of application critical data paths and improve the performance of the remote memory system.The memory tracking tool is used to emulate the memory access unit in the memory controller and build a prototype system.Experiment results show that the prefetching framework can achieve more than 90%accuracy and coverage.Compared with the Fastswap,the performance is improved by 59%.Compared with the default replacement framework of the kernel,the replacement framework improves application performance by 30%.

关 键 词：远端内存系统 软硬件协同 预取框架 替换框架 

分 类 号：TP316[自动化与计算机技术—计算机软件与理论] TP333[自动化与计算机技术—计算机科学与技术]