Image parallel ray tracing using static load balancing and data prefetching

Abstract

We propose an image parallel ray tracing with a static load balancing and data prefetching scheme on distributed memory multicomputers. We use a scattered decomposition of which the allocation unit is one pixel and each processor takes part in pixels scattered around the whole image space, which can make a load balance efficiently. In a scattered decomposition scheme, remote data accesses may be frequent since it can not utilize ray coherence sufficiently. To reduce the communication overhead due to this, we propose data prefetching by multicasting (DPM) where requested object data are also transferred to the processors dealing with the adjacent pixels if a processor requests a remote data. We implemented our approach on the Intel's Paragon and showed the linear speedup. However, the advantages of data prefetching did not materialize due to the machine's inefficient multicast operation.