Performance analysis of checkpointing strategies
Asser N. Tantawi, Manfred Ruschitzka · ACM Transactions on Computer Systems · 1984
A widely used error recovery technique in database systems is the rollback and recovery technique.Former models of rollback and recovery assumed Poisson failures and fixed (or exponential) checkpointing intervals.Extending these models, we consider general failure distributions.We also allow checkpointing intervals to depend on the reprocessing time and the failure distribution.Furthermore, failures may occur during checkpointing and error recovery.After deriving a general expression for system availability, we find that system availability resulting from the well-known equidistant checkpointing strategy depends only on the mean of the failure distribution.We then introduce a failure-dependent reprocessing-independent checkpointing strategy called equicost strategy.For Weibull failure distributions, which are good approximations of actual failure distributions of computer systems, we show that the equicost strategy achieves higher system availabilty than the equidistant strategy, which is known to be optimal under Poisson failures.