Roll-Forward and Rollback Recovery: Performance-Reliability Trade-Off

Trade-O Dhiraj K. Pradhan Nitin H. Vaidya Department of Computer Science Texas A&M University College Station, TX 77843-3112 fpradhan,[email protected] Abstract Performance and reliability achieved by a modular redundant system depend on the recovery scheme used. Typically, gain in performance using comparable resources results in reduced reliability. Several highperformance computers are noted for small mean time to failure. Performance is measured here in terms of mean and variance of the task completion time, reliability being a task-based measure de ned as the probability that a task is completed correctly. Two roll-forward schemes are compared with two rollback schemes for achieving recovery in duplex systems. The roll-forward schemes discussed here are based on a roll-forward checkpointing concept proposed in [5-8]. Roll-forward recovery schemes achieve signi cantly better performance than rollback schemes by avoiding rollback in most common fault scenarios. It is shown that the roll-forward schemes improve performance with only a small loss in reliability as compared to rollback schemes.