Decomposition Independence in Parallel Programs

Abstract

It is very common for parallel programs to give slightly different results when running on different numbers of processors or even when using the same number of processors but different data decompositions. Though this variation is usually not at a significant level it can mask programming errors by providing an alternative explanation for varying results. The obvious methods for solving this problem rely on performing the affected sections of the code in serial. This approach may require very large amounts of memory on the processor performing the serial calculation and may significantly extend the run-time of the program to an extent that is unacceptable even for debugging purposes.