基于并发性发掘的低开销回卷恢复实现方法-计算机工程.PDF

39 11 2013 11 Vol.39 No.11 Computer Engineering November 20 13 ·· 2013 A TP391 a a b ( a. b. 410082) 3 NAS NPB2.3 0.63 s3.19 s1.21 s 0.18 s0.67 s0.19 s 13.4%3.5%18.3%0.7%0.1%1.0% Implementation Method of Low Overhead Rollback Recovery Based on Concurrency Exploiting a a b YUAN Gong-biao , YANG Jin-min , BAI Shu-ren (a. College of Information Science and Engineering; b. Supercomputing Center, Hunan University, Changsha 410082, China) Abstract 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