Issue No.01 - January-March (1998 vol.6)
DOI Bookmark: http://doi.ieeecomputersociety.org/10.1109/4434.656777
The authors provide a detailed analysis of the memory-hierarchy effects in shared-memory architectures of one method of volume rendering-ray casting. They studied two parallel-partitioning and dynamic load-balancing algorithms-one object partition and one image partition-exploring trade-offs between their memory-hierarchy performance and the algorithmic optimizations they allow. Their resulting implementations (along with careful tuning of the ray-advancement kernel for Silicon Graphics' R8000) yield extremely high performance. For a 1-Gbyte female human-body data set, they attain an average frame rate of 1.0 frame per second, at a resolution of 400 pixels x 300 pixels, on a 16-processor Silicon Graphics Power Challenge. This is faster than the literature has previously reported for a data set this large. They have also extended their methods to a cluster of such machines. Using eight Silicon Graphics Power Challenge machines with a total of 64 processors, they attain average frame rates up to 10 frames per second on a 357-Mbyte male human-body data set for a sequence of frames generated by interactive user control.
Michael E. Palmer, Brian Totty, Stephen Taylor, "Ray Casting on Shared-Memory Architectures: Memory-Hierarchy Considerations in Volume Rendering", IEEE Concurrency (out of print), vol.6, no. 1, pp. 20-35, January-March 1998, doi:10.1109/4434.656777