While desktop grids are attractive platforms for executing parallel applications, their volatile nature has often limited their use to so-called “high-throughput ” applications. Checkpointing techniques can enable a broader class of applications. Unfortunately, a volatile host can delay the entire execution for a long period of time. Allocating redundant copies of each task to hosts can alleviate this problem by increasing the likelihood that at least one instance of each application task completes successfully. In this paper we demonstrate that it is possible to use statistical characterizations of host availability to make sound task replication decisions. We find that strategies that exploit such statistical characterizations are effective when compared to alternate approaches. We show that this result holds for real-world host availability data, in spite of only imperfect statistical characterizations. 1