wm@ogicse.ogi.edu (Wm Leler) (04/13/90)
The ParaGraph lecture series on Parallelism and Computer Graphics presents:
ISSUES AND ALGORITHMIC TECHNIQUES FOR PARALLEL COMPUTER GRAPHICS VISUALIZATION
Scott Whitman
Computer & Information Science Department
The Ohio Supercomputer Graphics Project
The Ohio State University
Commercial multiprocessors can be useful machines for generating
complex computer synthesized imagery quickly. There are a number of
issues involved in doing so that one does not typically encounter when
programming a von Neumann architecture. Among these are: exploiting
coherence in the image in a parallel context, handling of very large
databases, communication, load balancing, memory referencing, and
architectural characteristics. We have investigated a number of image
space based parallel approaches to the visualization problem for
implementation on a heirarchical shared memory multiprocessor in
particular. However, our results are applicable to all scalable
distributed memory MIMD architectures.
The approaches used can be partitioned according to a number of
different levels of granularity ranging from the pixel level to a frame
for an animation. In order to determine the most viable approach for
high performance, the algorithms were varied in their decomposition of
the problem, locality of data, and mapping of tasks to processors.
Testing of these algorithms was performed on the BBN Butterfly GP1000
and the new generation BBN TC2000. The real-world performance
indicates that high performance can be obtained when a graphics
algorithm is extensively analyzed and modified to exploit parallelism.
The analysis/modification/testing phase is typically longer for a
multiprocessor than a conventional architecture, but the benefit of
increased speed outweighs this extra cost.
The Speaker:
Scott Whitman is a Ph.D. candidate in the Computer and Information
Science Department at The Ohio State University. He received his M.S.
from Ohio State in Computer Science and his B.S. in Mathematics from
Carnegie-Mellon University. Previously, Scott worked at Cranston-Csuri
Productions as a Computer Graphics Researcher and at Evans & Sutherland
as an Applications Engineer. He organized the highly successful course
on Computer Graphics and Parallelism at the 1989 SIGGRAPH conference,
and will be heading up the same course this year. Scott's research
interests include: Image Synthesis, Advanced Computer Architectures,
and Parallel Programming.
Wednesday, April 18, 4:00 pm
At OCATE (building E-3 in the Oregon
Graduate Institute Science Park)
19500 NW Gibbs Drive
Beaverton,OR 97006
503/690-1460