Another Two-Level Failure Recovery Scheme : Performance Impact of Checkpoint Placement andCheckpoint Latency

This report deals with the design and evaluation of a \two-level" failure recovery scheme for distributed systems. In our previous work 30, 32], we motivated a \two-level" recovery approach that tolerates the more probable failures with a low overhead, and less probable failures with possibly higher overhead. The two-level approach can achieve a smaller overhead as compared to traditional recovery schemes. The contributions of this report are summarized below: We present and evaluate a \two-level" recovery scheme that is suitable for a network of workstations, each workstation having a local disk. The recovery scheme presented in the report can tolerate transient processor failures with a low overhead , while other failures require a larger overhead. The report presents analysis of the average (expected) task completion time using the proposed scheme. This scheme has been implemented on a workstation cluster. Our analysis indicates that the proposed two-level recovery scheme can achieve better performance as compared to existing \one-level" recovery schemes. The report also evaluates the impact of checkpoint latency on the performance of the recovery scheme. To our knowledge, no analysis of the performance impact of checkpoint latency has been carried out previously. Experimental measurements of checkpoint latency and checkpoint overhead for four applications are presented. References 32, 30] present material related to this report. The interested reader can obtain these references via anonymous ftp from ftp.cs.tamu.edu:/pub/vaidya. y This report was revised several times in January 1995. The purpose of these revisions was to add Sections 10 and 11, and to revise Section 1.