超级计算系统互连网络带内管理的实现与评测  

The Implementation and Evaluation of In-Band Network Management in Supercomputing System

作　　者：曹继军[1] 肖立权[1] 王克非[1] 庞征斌[2] 陈琳[1] 

机构地区：[1]国防科学技术大学计算机学院,长沙410073 [2]国防科学技术大学并行与分布式计算重点实验室,长沙410073

出　　处：《计算机学报》2016年第9期1717-1732,共16页Chinese Journal of Computers

基　　金：国家"八六三"高技术研究发展计划项目基金(2012AA01A301;2013AA014301;2013AA01A208)资助~~

摘　　要：互连网络是超级计算系统的关键部件,其易管理性将直接影响整个系统的RAS特性(可靠性、可用性和服务性)."天河二号"超级计算系统采用了定制的高速互连网络,该网络由大约5856块网络交换芯片NRC和18 304块网络接口芯片NIC构成,其互连规模极其巨大.为了实现对系统内众多网络芯片及其端口的高效率配置和监控等管理操作,该互连网络采用了基于带内的网络管理技术.该文描述了带内网络管理的设计与实现,具体包括网络管理功能、网络管理总体结构、带内管理描述符格式与报文处理流程、管理软件的总体框架等.基于部署在国家超级计算广州中心的实际系统,该文对"天河二号"互连网络带内的管理进行了测试,测试结果证明了带内管理的高效性.Interconnect network plays an important role in super computing system. And its manageability directly affects the RAS （Reliability, Availability and Serviceability） characteristics of the whole system. The Tianhe-2 supercomputing system uses proprietary high-speed interconnect network, which includes 5856 high-radix network router chips （NRC） and 18 304 network interface chips （NIC）. For the very large-scale interconnect network, it is a great challenge to manage （such as configure and monitor） the numerous network chips and its ports in an efficient way. By implementing the in-band management scheme,we achieve a very efficient network management for the interconnect network in Tianhe-2 system. In this paper, we introduced the design of in-band network management in Tianhe-2 system, especially emphasizing on several key features, which include the basic functionalities and architecture of network management, format of management descriptors, dataflow and processing of management packets, basic frame of management software, etc. In this paper, we also evaluated the performance of in-band network management on the supercomputer that located in NSCC-GZ. The results demonstrate the efficiency of the in-band management for interconnect network.

关 键 词：超级计算系统 互连网络 带内管理 天河二号 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]