高性能并行计算系统检查点技术的应用.pdfVIP

高性能并行计算系统检查点技术与应用 孙国忠 李艳红 樊建平 (中国科学院计算技术研究所 中国科学院研究生院 北京 100080) (sgz@ncic.ac.cn,lyh@ncic.ac.cn,fan@ict.ac.cn) 摘 要 随着高性能并行计算系统规模越来越大，软件和硬件发生故障的概率随之增大，系统的容错性和可靠性已经 成为应用可扩展性的主要限制因素。并行检查点技术可以使系统从故障中恢复并减少计算损失，是高性能计算系统重 要的容错手段。本文将介绍检查点技术的背景和定义，研究并行检查点协议的分类，检查点存储技术，以及利用这些 协议和技术实现的MPI并行检查点系统，最后给出对各个关键技术的详细评价及结论。 关键词 高性能计算;消息传递系统;并行检查点;回滚恢复 中图法分类号 TP31 A Survey of Checkpointing Technology and It’s Application for High Performance Parallel Systems Sun Guo-Zhong Li Yan-Hong Fan Jian-Ping (Institute of Computing Technology,Chinese Academy of Sciences/Graduate School of the Chinese Academy of Sciences, Beijing 100080 ) (sgz@ncic.ac.cn lyh@ncic.ac.cn fan@ict.ac.cn) Abstract With the scale of high performance parallel computing systems becoming larger,the fault probability of software and hardware in these systems is increased.As a result, issues of fault tolerance and reliability are becoming limiting factors on application scalability.Parallel checkpointing can help fault system recover from fault and reduce the computing losing,and is an important method for tolerating fault of high performance computing system.This paper will discuss the background and definitions of checkpointing,classify of parallel checkpointing protocols, checkpoint storage technology, and several MPI systems adopting these parallel checkpointing protocols.At last we give appraisement of these key technologies and list our conclusions. Key words High Performance Computing; Message Passing System; Parallel Checkpointing ; Rollback Recovery 1 引 言 具有对硬件故障的容错技术。采用检查点技术恢 复应用运行是一种有效的容错方法。 高性能并行计算领域的容错技术由于以下几 检查点技术除了实现系统容错，还能协助实现 种情况而越发受到重视。1）在一台高性