Nested Parallelism with Algorithmic Skeletons

Abstract

New trend in design of computer architectures, from memory hierarchy design to grouping computing units in different hierarchical levels in CPUs, pushes developers toward algorithms that can exploit these hierarchical designs. This trend makes support of nested-parallelism an important feature for parallel programming models. It enables implementation of parallel programs that can then be mapped onto the system hierarchy. However, supporting nested-parallelism is not a trivial task due to complexity in spawning nested sections, destructing them and more importantly communication between these nested parallel sections. Structured parallel programming models are proven to be a good choice since while they hide the parallel programming complexities from the programmers, they allow programmers to customize the algorithm execution without going through radical changes to the other parts of the program. In this thesis, nested algorithm composition in the STAPL Skeleton Library (SSL) is presented, which uses a nested dataflow model as its internal representation. We show how a high level program specification using SSL allows for asynchronous computation and improved locality. We study both the specification and performance of the STAPL implementation of Kripke, a mini-app developed by Lawrence Livermore National Laboratory. Kripke has multiple levels of parallelism and a number of data layouts, making it an excellent test bed to exercise the effectiveness of a nested parallel programming approach. Performance results are provided for six different nesting orders of the benchmark under different degrees of nested-parallelism, demonstrating the flexibility and performance of nested algorithmic skeleton composition in STAPL .