Performance Evaluation Tools for Message-Passing Parallel Problems

Abstract

The article presents a number of issues of designing and implementing performance evolution tools for message-passing parallel programs, eg. MPI and PVM. There is at number ofspecial techniques for investigating parallel programs, whose implementations are tools presented, A concept of performance observability is introduced. Although a number of interesting solutions of performance tools was developed in the course of last decade, there is a great demand for portable and integrated tools. Understanding the reason for this situation requires evaluating the state of art of existing tools, their advantages and drawbacks. Due to a complicated mechanism of interactions between a tool and operating system, computer architecture and application, evaluating the tool includes taking into account a large number of features. There is introduced a set of criteria, which enable a thorough evaluation of tools. based on the work of HPC standardizing organizations as well as the author's work. The second part of the article presents the review of features of the particular tools developed over last decade. The tools are being evaluated on the base of the criteria introduced. The features of the PARNAS performance tool project and its implementation are presented. The summary presents further avenues of inquiry in parallel performance evaluation tools.

Keywords

• Distributed and grid systems