摘要
以非易失性存储器(NVM)作为主存且以动态随机存取存储器(DRAM)作为片外高速缓存(EC),是一种可以满足大数据应用内存容量需求的新型混合内存结构(ECNVM)。该结构同时具有NVM的大存储容量和DRAM的低存取延迟的优点。传统的结构是以片内SRAM作为片内高速缓存(IC)且以DRAM作为主存(IC-DRAM)。与ICDRAM相比,EC-NVM在容量比、延迟比方面均有显著不同,导致在IC-DRAM场景下的设计方法和优化策略直接迁移到EC-NVM上未必有良好的效果。本文评测了 EC-NVM的体系结构特性(包括高速缓存粒度、关联度、替换算法、预取算法等),获得了指导ECNVM结构的设计和优化的一系列发现。
To satisfy the memory capacity requirement of big-data applications,a promising architecture,in which the non-volatile memory(NVM)is used as main memory and dynamic random access memory(DRAM)is taken as external cache(EC),is proposed,referred to as EC-NVM.Compared to traditional memory architecture that involves the last level cache(LLC)and DRAM(referred to as IC-DRAM),EC-NVM is significantly different in terms of capacity and latency,thus traditional design cannot be directly borrowed without reconsideration.This paper explores the characteristics and architectural implications of EC-NVM in diverse dimensions including cache line granularity,associativity,replacement policy and prefetching methods.All findings provide valuable hints for the designers of EC-NVM.
作者
潘海洋
刘宇航
卢天越
陈明宇
Pan Haiyang;Liu Yuhang;Lu Tianyue;Chen Mingyu(Advanced Computing System Laboratory,Institute of Computing Technology,Beijing 100190;University of Chinese Academy of Sciences,Beijing 100049;Peng Cheng Laboratory,Shenzhen 518055)
出处
《高技术通讯》
CAS
2021年第5期464-478,共15页
Chinese High Technology Letters
基金
国家自然科学基金面上项目(61772497)
国家重点研发计划(2017YFB1001600)资助项目。