high performance computing on graphics processing units: hgpu.org

SAPFOR (System For Automated Parallelization) is a software development suite that is focused on cost reduction of manual program parallelization. SAPFOR produces parallel programs according to the high-level DVMH parallel programming model. SAPFOR relies on an implicitly parallel programming model, so it includes an automatic parallelizing compiler. On the other hand, it allows the user to guide parallelization. The system provides the user with a set of source-to-source transformations which can be performed in an automatic way. The user may also assert some implicit program properties useful for parallelization. This paper presents the interactive subsystem of SAPFOR and discusses how it supports iterative parallelization. We advocate the use of the client-server model to organize the interaction with the user. We also present the data transfer optimization technique the automatic parallelizing compiler implements. This technique reduces communication overhead when GPU is used to execute a parallel program. We guide SAPFOR to perform semi-automatic parallelization of the NAS Parallel Benchmarks 3.3.1. The paper evaluates the performance of parallel versions that SAPFOR builds automatically and compares it with the performance of manually written versions.