曙光5000A高效能计算节点的设计与实现  

Design and Realization of Dawning 5000A High-productivity Computing Node

作　　者：曾宇[1,2] 王洁[1,2] 孙凝晖[1] 

机构地区：[1]中国科学院计算技术研究所计算机系统结构重点实验室,北京100080 [2]中国科学院研究生院,北京100039

出　　处：《计算机工程》2009年第6期17-19,22,共4页Computer Engineering

基　　金：国家“863”计划基金资助重大项目(2006AA01A102)

摘　　要：由于求解问题和系统规模的不断扩大,基于cluster架构的高性能计算机面临扩展性、可靠性、功耗、占地面积、均衡性等诸多挑战。该文针对计算模块、交换管理模块、自适应功率管理、专用FPGA硬件加速部件、高速PCI-E全交换扩展等方面,设计并实现高效能计算节点。基于该节点构建的曙光5000A百万亿次计算机能有效解决计算密度、I/O扩展及带宽瓶颈和能耗等方面的瓶颈。As for the scale of the problem and the system continues to expand, the cluster-based high-performance computer is facing scalability, reliability, power consumption, footprint, balance, and many other challenges. This paper introduces the design and realization of high-productivity computing node such as computing module, switch module, management module, adaptive power management, FPGA-based hardware accelerator board, high-speed PCI-E switch extend module and other aspects. It resolves the computing density, I/O expansion and bandwidth bottleneck as well as energy consumption and other bottlenecks in Dawning 5000A 100 Teraflops supercomputer based on the high-productivity computing node.

关 键 词：高效能 节能 硬件加速 

分 类 号：TP303[自动化与计算机技术—计算机系统结构]