Efficient garbage collection schemes for causal message logging with independent checkpointing in message passing systems
JinHo Ahn, Chong‐Sun Hwang · 2005
Causal message logging approach with independent checkpointing having two advantages, low failure-free overhead and limited rollback, is very attractive for providing fault-tolerance for parallel/distributed applications on message passing systems. However, to satisfy the two desirable properties, this approach should force each process to maintain a large amount of message logging information in its volatile storage and piggyback a part of the information on each sending message. In this paper, we present two efficient garbage collection schemes for avoiding these drawbacks. The first scheme enables each process to remove log information related to all messages no longer needed for process recovery by forcing the process to piggyback only an additional -size vector, where is the number of processes, on each sending message without requiring any additional message. The second scheme allows the process to remove a part of the rest of log information in storage buffers if more empty buffer space is required after the first scheme has been executed. The two schemes considerably reduce the number of additional messages and forced checkpoints needed by the garbage collection and output commit overhead compared with previous schemes.