Capsules: Expressing Composable Computations in a Parallel Programming Model

Abstract

A well-known problem in designing high-level parallel programming models and languages is the "granularity problem", where the execution of parallel task instances that are too fine-grain incur large overheads in the parallel runtime and decrease the speed-up achieved by parallel execution. On the other hand, tasks that are too coarse-grain create load-imbalance and do not adequately utilize the parallel machine. In this work we attempt to address this issue with a concept of expressing "composable computations" in a parallel programming model called "Capsules". Such composability allows adjustment of execution granularity at run-time. In Capsules, we provide a unifying framework that allows composition and adjustment of granularity for both data and computation over iteration space and computation space. We show that this concept not only allows the user to express the decision on granularity of execution, but also the decision on the granularity of garbage collection, and other features that may be supported by the programming model. We argue that this adaptability of execution granularity leads to efficient parallel execution by matching the available application concurrency to the available hardware concurrency, thereby reducing parallelization overhead. By matching, we refer to creating coarse-grain Computation Capsules, that encompass multiple instances of fine-grain computation instances. In effect, creating coarse-grain computations reduces overhead by simply reducing the number of parallel computations. This leads to: (1) Reduced synchronization cost such as for blocked searches in shared data-structures; (2) Reduced distribution and scheduling cost for parallel computation instances; and (3) Reduced book-keeping cost maintain data-structures such as for unfulfilled data requests. Capsules builds on our prior work, TStreams, a data-flow oriented parallel programming framework. Our results on an SMP machine using the Cascade Face Detector, and the Stereo Vision Depth applications show that adjusting execution granularity through profiling helps determine optimal coarse-grain serial execution granularity, reduces parallelization overhead and yields maximum application performance.

A NUMA-Aware Fine Grain Parallelization Framework for Multi-core Architecture

“…Another possible approach, as given by Capsules [16], requires the user to define several grain sizes. The runtime then chooses which grain best matches the current situation.…”