The static load balancing strategy used in the current implementation has been successfull on tests run to date. It achieves better than 92% efficiency in tests with up to 72 nodes on an Intel Paragon for the "balls" scene from the SPD package. The iPSC/860 and the Paragon both benefit from the concurrent I/O facilities (CFS and PFS) they have. The concurrent I/O performance advantage is quite noticable when rendering animations. Workstation clusters and the SP-2 suffer from the lack of parallel I/O. Parallel I/O helps the scalability of the rendering library significantly.
Performance testing has been conducted using the SPD database maintained by Eric Haines . Table 1 shows the timing results for the "balls", "tetra", and "tree" scenes when run on a cluster of Suns and an Intel iPSC/860. The Suns used in the SPD timings were SparcStation 10 model 40 workstations running Solaris 2.3. The SPD results shown were performed according to the SPD guidelines, and include spatial subdivision preprocessing time as well as ray tracing time.
The SPD results indicate that there is substantial room for improvement in the spatial subdivision implementation currently used. Results from previous tests show that only 75% to 90% of the ray intersection tests are being eliminated by the current subdivision scheme. Any improvement in the spatial subdivision performance will translate directly into a corresponding decrease in run time both sequentially and in parallel.