Performance, Scalability, and Robustness in the Harness Metacomputing Framework

Abstract

Harness is a software framework and methodology that facilitates distributed metacomputing via a reconfigurable plugin-based architecture. The current version of Harness emulates multiple concurrent programming environments including PVM, MPI, JavaSpaces, and generic task parallelism models. Supporting multiple parallel programming models within a unified framework has high utility and flexibility, and is capable of evolving as needed. Harness, a modular software system, is able to deliver such functionality and support legacy as well as new programming paradigms without loss of efficiency. In a set of benchmarking experiments, the performance of applications executed using the Harness-PVM system was within a few percent of their performance using native PVM. Furthermore, the Harness-PVM system scales well and exhibits increased tolerance to failures, due to its event-based distributed architecture. Details of the Harness infrastructure, features that contributed to enhanced scalability and failure-resilience, and performance results are presented.