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(FFF)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.