Support for parallel out of core applications on Beowulf workstations

Beowulf workstations are a new approach to parallel computing that combine the characteristics of the Pile-Of-PC architecture with freely available software and provide high performance at a low cost for many applications. In order to provide this performance in the face of increasing problem sizes, efficient out of core (OOC) methods must be investigated. In light of shortcomings in operating system virtual memory performance, we have implemented a parallel file system and a new interface tailored to OOC algorithms to provide efficient and convenient access to disk resources in Beowulf workstations. This paper focuses on how these components extend the range of problem sizes which can be efficiently solved on Beowulf workstations in the context of a sample application, a Gauss-Seidel iterative solver. Execution times of the solver when using traditional virtual memory and the parallel file system are compared to characterize performance for various problem sizes, and it is concluded that the combination of parallel file system and new interface significantly increases the size of problems that can be efficiently solved.