In order to achieve real-time performance, one is faced with several issues related to machine and algorithm limitations. Only highly balanced systems can hope to achieve maximum performance at a reasonable cost. In order to analyze the performance of PVR, we must have some idea of the performance of every component in our system. Our goal is to understand (and possibly predict) the rendering time of each image.
This study is based on the Intel Paragon. The Paragon uses a multitude of nodes each of which has two processors, the Intel i860XPs. Using the OSF/1 operating system, only one of the processors is available for computation. We have used the machine at Stony Brook, which has 96 users' nodes (with 32MB of RAM each) and runs OSF/1. Only a portion of the memory is available for the application.