Interactive parallel rendering on a multiprocessor system with intelligent communication controllers

Most data-parallel rendering algorithms (on multiprocessot systems with distributed memory) spend a substantial amount of time composing (merging or assembling) the partial images of all the processors. This paper shows how "intelligent communication controllers" (ICCs) help to reduce the immense communication overhead and accumulated latencies to an absolute minimum. Three examples of "intelligent" communication schemes are presented: the fully automatic redistribution of multi-dimensional data sets, depthmerge and bucket-sort. We show that these (and other) "intelligent communication schemes" can be implemented in hardware with a reasonable effort and that the communication bandwidth is used most efficiently. This results in a good speed-up, good scalability and the maximum utilizable performance for parallel rendering and many other data-parallel algorithms running on our multiprocessor system "MUSIC". As an example, we present a simple objectparallel renderer running at interactive frame rates.