Performance evaluation of a graphic accelerators cluster

High performance computer systems based on commercial off the shelf components, namely Beowulf class systems, have been widely adopted to develop and test scientific and engineering applications. Each computing node of such distributed memory multicomputer is a PC or a workstation connected to the others within a system area network. In this paper we show how to harness the computing power of accelerated graphic cards available in each node and realizing an efficient, portable and scalable software library for parallel hardware rendering. A detailed performance analysis of a graphic accelerators cluster, based on the evaluation of a distributed memory VRML (Virtual Reality Modeling Language) browser, is given in terms of execution times and well known performance parameters.