We have shown that using PARC cubes for measuring useful work generates an intuitive way to load balance volume ray casting on distributed memory parallel machines. This not only generates a method that is theoretically sound but its preliminary implementation seems to present a method that is both efficient and scalable.
We have also proposed a new method for compositing that achieves better throughput than previous methods and that can be used to generate better refresh rates. If one cannot accept the delay pipelining imposes, one can always make judicious replication of volume data, for instance, one volume for every 16 processors to avoid long image delay times and still keep high refresh rates.
We believe our method is simple, fast, uses coherency and achieves high resource utilization on a given machine. As we use PARC, we achieve a high utilization of the compute processors and thus a very fast rendering time on every processor. Because of our pipelined compositing scheme, we achieve a much higher network utilization than other methods. Finally, our feedback synchronization image request technique guarantees a constant flow of information that adapts itself to different configurations of processor performance and network utilization.
Our current implementation can be greatly improved and optimized. One of our main concerns is to smoothly integrate all the parallel code into VolVis, so our users can take advantage not only of its intuitive and flexible user interface, but also of greater speed provided by parallel machines. Other plans include the porting of our algorithm to other architectures and a more detailed performance analysis of the whole algorithm. We are also planning on introducing optimization that would allow the system to use data replication and sharing whenever allowed. This way users with multiple processor shared-memory machines, like a network of Sparc1000s would be able to get better performance.
Another direction of future work is the extension of our load balancing technique to non-slabs partitions. The major problem is that computing optimal partitions in one dimention (the slab case) is already hard and computationally expensive. Another interesting question is whether this method can be extended to irregular shaped grids.