News

Joint Seminar, Computer Science and Electrical, Computer and Systems Engineering

Finding the Right Type of Parallelism in an "Embarrassingly Parallel" Algorithm

Abe Stephens
University of Utah

Wednesday, June 20, 2007

Abstract:

Unlike many areas of computer science interactive ray tracing algorithms have evolved several times to remain at the cutting edge of computing performance. Early work on multiprocessor super computers directed the field towards fine-grained task-parallelism. Later as desktop performance increased and memory became the predominant bottleneck, algorithms evolved to increase memory coherence by operating more coarsely on packets of data in parallel. Today these algorithms are undergoing another transition to utilize wide data parallel programming models on GPUs. These processors rely on extensive SIMD operational coherence. This talk will describe interactive ray tracing's transition from massive task parallel super computers to multi-core CPUs, and recently to high performance commodity data parallel models like CUDA. This transition illustrates how other workloads might be adapted to leverage increasing degrees of parallelism in mainstream platforms.

Bio: Abe Stephens is a PhD student at the University of Utah's Scientific Computing and Imaging Institute where he works with Steven Parker. His research focuses on parallel techniques for temporally adaptive rendering and large data visualization. He is a principle contributor to the Manta Interactive Ray Tracer and has worked with Intel and Silicon Graphics to improve interactive ray tracing techniques on parallel systems. Abe has published several papers on interactive ray tracing and has spoken at Siggraph and Eurographics courses on the subject. He received a BS in Computer Science from Rensselaer Polytechnic Institute in 2003.

Hosted by: Carlos A. Varela (x6912)

Last updated: June 8, 2007