Optimizations on Array Skeletons in a Shared Memory Environment

Abstract

Map- and fold-like skeletons are a suitable abstractions to guide parallel program execution in functional array processing. However, when it comes to achieving high performance, it turns out that confining compilation efforts to individual skeletons is insufficient. This paper proposes compilation schemes which aim at reducing runtime overhead due to communication and synchronization by embedding multiple array skeletons within a so-called spmd meta skeleton. Whereas the meta skeleton exclusively takes responsibility for the organization of parallel program execution, the original array skeletons are focussed to their individual numerical operation. While concrete compilation schemes assume multithreading in a shared memory environment as underlying execution model, ideas can be carried over to other settings straightforwardly. Preliminary performance investigations help to quantify potential benefits.KeywordsData DependencyParallel ExecutionProgram ExecutionNumerical OperationRuntime OverheadThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.