面向大数据的异构内存系统  

Big data oriented hybrid memory systems

作　　者：王孝远 廖小飞[1] 刘海坤[1] 金海[1] WANG Xiaoyuan;LIAO Xiaofei;LIU Haikun;JIN Hai(School of Computer Science and Technology,Huazhong University of Science and Technolog)

机构地区：[1]华中科技大学计算机学院,湖北武汉430074

出　　处：《大数据》2018年第4期15-34,共20页Big Data Research

基　　金：国家重点研发计划基金资助项目(No.2017YFB1001603);国家自然科学基金资助项目(No.61672251;No.61732010;No.61628204)~~

摘　　要：受限于DRAM和新型非易失性存储器(non-volatile memory,NVM)的缺陷,单纯的DRAM或者NVM难以满足大数据应用对内存系统容量以及功耗提出的高要求。因此如何将DRAM和NVM组合成异构内存系统并进行高效的管理、准确的评估,是当今学术界和工业界面临的主要挑战。从体系结构、系统软件、编程模型以及应用等方面对面向大数据的异构内存系统进行分析与研究,提出了一系列异构内存系统的优化方法,如层次化异构内存架、片上缓存管理、访存调度、能耗管理、虚实地址转换和面向对象的内存分配与迁移机制等,并实现了原型系统进行验证。Big data applications put more pressure on memory system in the aspect of capacity and power consumption. However, limited by the shortcomings of DRAM and non-volatile memory（NVM）, memory consists of single medium like DRAM or NVM is not competent for the requirements of big data applications. Thus, how to effectively design, efficiently manage and accurately evaluate the hybrid memories consist of DRAM and NVM are the major challenges that the academia and industry face today. The challenges of hybrid memory systems for big data processing from the perspective of computer architecture, system software, programming model and application were analyzed, and several solutions and optimizations were correspondingly provided, such as on-chip cache management, parallel processing, memory access scheduling, energy management, virtual-to-physical address translation, object-level memory allocation and migration mechanisms. Meanwhile, a number of prototypes to validate the effectiveness and efficiency of these proposals were developed.

关 键 词：内存计算 异构内存 大数据 非易失性存储器 

分 类 号：TP31[自动化与计算机技术—计算机软件与理论]