An aperiodic checkpointing strategy in desktop grids

In the context of desktop grids, resource failures occur frequently and checkpointing is helpful in accelerating the completion of long-running jobs. Since failure distributions of individual hosts in desktop grids are diverse and are hard to model, checkpointing strategies based on a closed form of failure distribution are not good choices for such cases. We propose an aperiodic checkpointing strategy which dynamically sets checkpoint according to history information of interfailure lengths. This strategy works for each individual host based on a sample of its history availability interval lengths. Rather than deriving a closed form for the probability distribution of availability interval lengths, the strategy directly approximates it with the sample. We conduct trace driven simulations to compare this strategy and periodic strategy, and results testify its effectiveness. Besides an unsuccessful attempt to improve the strategy, a requirement of using the strategy is detailed in this paper.