Computer Science

Colloquium Details

A Region-based Approach to Parallel Computing

Abstract

In spite of continual advances in hardware speed and sophistication, parallel computers remain a somewhat underutilized tool due to the challenges that users face when it comes to implementing parallel programs. The primary approaches that are in active use today are very low-level, requiring great effort by programmers and forcing them to manage the details of parallel programming manually. Other approaches support a higher degree of abstraction, yet make it challenging for users to predict the effects that their algorithmic choices will have on the implementation. Both the high-A Region-based Approach to Parallel Computing level and low-level approaches present significant obstacles toward portable performance.

ZPL is a parallel programming language that strives to bridge this gap by supporting a high-level syntax, yet one that permits programmers to understand the parallel implementation of their code at the source level. The key to this approach has been the "region" -- a language-based representation of an index set. In this talk, I will give a brief introduction to regions and their use in the ZPL language. I will also summarize the benefits that regions provide, both to the user and to the compiler. I will then describe my recent experiments that compare a ZPL implementation of the NAS MG benchmark to versions written using other current approaches to parallel programming: MPI, High Performance Fortran (HPF), Co-Array Fortran, and Single Assignment C. Each approach is evaluated in terms of its performance, its portability, and its ability to express the algorithm clearly. Our results show that ZPL does well in all three areas. As time permits, I will give a very brief summary of my current work, which is to extend ZPL to support parallel sparse computation. This work will be described in greater detail at a seminar to be held on March 2nd.