基于并发性发掘的低开销回卷恢复实现方法  

Implementation Method of Low Overhead Rollback Recovery Based on Concurrency Exploiting

作　　者：袁功彪[1] 杨金民[1] 白树仁[2] 

机构地区：[1]湖南大学信息科学与工程学院,长沙410082 [2]湖南大学超级计算中心,长沙410082

出　　处：《计算机工程》2013年第11期46-51,共6页Computer Engineering

基　　金：国家自然科学基金资助项目(61272401;61133005);湖南省科技计划基金资助重点项目(201GK2003)

摘　　要：现有的回卷恢复容错技术存在同步约束和阻塞问题,其时间开销随系统节点规模的增大而剧增。为此,提出一种基于并发性发掘的低开销回卷恢复实现方法。利用消息传递附带跟踪消息依赖的策略解除消息日志中的同步约束,解析进程负载以发掘进程负载的并发性,构建进程负载并发执行的实现架构,采用数据缓存策略和多线程技术实现进程内部各负载的并发执行,以降低故障恢复开销。3个NAS NPB2.3标准性能检测程序的实验结果表明,该方法可使检查点开销从0.63 s、3.19 s、1.21 s分别降低到0.18 s、0.67 s、0.19 s,日志开销率从13.4%、3.5%、18.3%分别降低到0.7%、0.1%、1.0%。Existing rollback recovery technologies show that their time overheads increase sharply with the scale of nodes due to synchronization constraints and the sequential execution pattern. Aiming at this problem, this paper proposes an implementation method of low overhead rollback recovery based on concurrency exploiting. It uses the strategy of piggybacking dependency on messages to release the message log synchronization constraints. In addition, the workloads in a process is resolved to exploit their concurrency. Then data buffering strategy and multithreading technology are applied to implement the concurrent execution of various process workloads, leading to a low overhead rollback recovery scheme. Experimental results of three NAS NPB2.3 benchmarks show that the overheads of a checkpoint are decreased from 0.63 s, 3.19 s, 1.21 s to 0.18 s, 0.67 s, 0.19 s respectively, and the overhead ratios of message logging are decreased from 13.4%, 3.5%, 18.3% to 0.7%, 0.1%, 1.0% respectively.

关 键 词：回卷恢复 时间开销 同步约束 并发性挖掘 消息日志 检查点 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]