Probabilistic Allocation of Synchronous Tasks to Desktop Grids Resources

Abstract

While desktop grids are attractive platforms for executing parallelapplications, their volatile nature has often limited their use toso-called {"}high-throughput{"} applications. Checkpointing techniques,suchas used in MPICH-V, can enable a broader class of applications.Unfortunately, a volatile host can delay the entire execution for a longperiod of time. Allocating redundant copies of each task to resources canalleviate this problem by increasing the likelihood that at least oneinstance of each application task complete successfully. In this paperwedemonstrate how it is possible to use statistical characterizations ofhostavailability to make sound task replication decisions. We findthat strategies that exploit such statistical characterizations areeffective when compared to alternate approaches. Perhaps moreimportantly,we show that this result holds for real-world host availability data forwhich only imperfect statistical characterizations are possible.