Exploiting Monotone Convergence Functions in Parallel Programs

Abstract

Scientific codes which use iterative methods are often difficult to parallelize well. Such codes usually contain \code{while} loops which iterate until they converge upon the solution. Problems arise since the number of iterations cannot be determined at compile time, and tests for termination usually require a global reduction and an associated barrier. We present a method which allows us avoid performing global barriers and exploit pipelined parallelism when processors can detect non-convergence from local information. (Also cross-referenced as UMIACS-TR-96-31.1)