Latency and Overhead in Transparent Distributed Rollback- Recovery

点击翻译按钮获取中文摘要

This research investigates the tradeoff between failure-free overhead and output latency in transparent distributed rollback-recovery. This tradeoff arises because, with rollback-recovery, output to the "outside world" must be delayed until enough recovery information is recorded on stable storage to guarantee that if a failure occurs, the system will never roll back to a state preceding the one from which the output was performed. The resulting latency can be reduced by keeping recovery information on stable storage closely up-to-date during failure-free operation, but at the expense of extra failure- free overhead. The work includes an investigation of how rollback-recovery methods based on logging can be extended to support nondeterministic process execution, occurring, for instance, as a result of multithreading. Nondeterministic events are recorded on stable storage, and are used during failure recovery to replay the same events. For example, by recording the execution of synchronization primitives such as semaphores, one can support multithreading in which all access to shared data is protected by these synchronization primitives. Checkpoints offer an alternative for processes for which recording nondeterministic events is impossible or is too expensive.