Runtime support for virtual BSP computer

Abstract

Several computing environments including wide area networks and nondedicated networks of workstations are characterized by frequent unavailability of the participating machines. Parallel computations, with interdependencies among their component processes, can not make progress if some of the participating machines become unavailable during the computation. As a result, to deliver acceptable performance, the set of participating processors must be dynamically adjusted following the changes in computing environment. In this paper, we discuss the design of a run time system to support a Virtual BSP Computer that allows BSP programmers to treat a network of transient processors as a dedicated network. The Virtual BSP Computer enables parallel applications to remove computations from processors that become unavailable and thereby adapt to the changing computing environment. The run time system, which we refer to as adaptive replication system (ARS), uses replication of data and computations to keep current a mapping of a set of virtual processors to a subset of the available machines. ARS has been implemented and integrated with a message passing library for the Bulk-Synchronous Parallel (BSP) model. The extended library has been applied to two parallel applications with the aim of using idle machines in a network of workstations (NOW) for parallel computations. We present the performance results of ARS for these applications.

This work was partially supported by NSF Grant CCR-9527151. The content does not necessarily reflect the position or policy of the U.S. Government.