Exploiting monotone convergence functions in parallel programs

Scientific codes which use iterative methods are often difficult to parallelize well. Such codesusually contain while loops which iterate until they converge upon the solution. Problems arisesince the number of iterations cannot be determined at compile time, and tests for terminationusually require a global reduction and an associated barrier. We present a method which allowsus avoid performing global barriers and exploit pipelined parallelism when processors can detect...