面向云计算的多核处理器存储和网络子系统优化设计  

Optimization of memory and network subsystems in multi-core processors for cloud computing

作　　者：苏文[1,2,3,4] 王焕东[1,2,4] 台运方[1,2,3,4] 王靖[1,2,3,4] 

机构地区：[1]中国科学院计算机系统结构重点实验室,北京100190 [2]中国科学院计算技术研究所,北京100190 [3]中国科学院研究生院,北京100049 [4]北京龙芯中科技术服务中心有限公司,北京100190

出　　处：《高技术通讯》2013年第4期360-367,共8页Chinese High Technology Letters

基　　金：国家"核高基"科技重大专项(2010ZX01036-001-002;2009ZX01028-002-003;2009ZX01029-001-003);国家自然科学基金(61100163;61133004)资助项目

摘　　要：针对传统多核处理器设计缺乏对虚拟机和典型云服务的支持的问题,分析了云计算数据中心和虚拟机的基本架构和特点,指出数据传输、网络性能、I/O虚拟化是一直未被关注的影响系统性能的关键因素,进而提出了一种改进的多核处理器设计方案。该方案通过采用片上内存拷贝引擎、改进直接内存访问(DMA)设计、改进直接缓存访问(DCA)设计和采用快速地址转换和远程内存访问(RDMA)技术,来较大地提高存储系统、网络、I/O的性能和系统的并行性。实验表明,该方案实现的单核800MHz处理器千兆以太网络TCP传输带宽较传统方案提高48.2%并达到峰值800Mbps,内存拷贝操作加速比达到14倍以上,快速傅立叶变换(FFT)和矩阵乘法加速比达到2倍以上,同时系统高速缓存效率显著提高。To solve the problem that traditional multi-core processer designs rarely consider to give support to virtual ma- chines and typical cloud services, the paper analyzes the typical architecture and characteristics of data-centers for cloud computing and virtual machines and points out that data transfer, network performance and I/O virtualization are the key but neglected factors affecting system performance, and based on the analysis, proposes an improved multi-core processor design scheme. The scheme uses the innovative technologies of on-chip memory copy engine, improved direct memory access （DMA）, improved direct cache access （DCA）, fast address translation and remote DMA （RDMA） to increase the performance of memory system, network, I/O and parallelism. The experimental results show that when the scheme is used, the TCP bandwidth can be increased by 48.2%, a peak bandwidth 800Mbps can be achieved in a Gigabit Ethernet, the memory copy can reach a speed up of 14 times compared with the traditional methods, the speeds of fast fourier transform （FFT） and matrix multiplication can be doubled, while the cache efficiency is significantly improved.

关 键 词：云计算 多核处理器 网络优化 虚拟机 计算机体系结构 

分 类 号：TP332[自动化与计算机技术—计算机系统结构]