Distributed rendering of interactive soft shadows
Michael Isard, Mark Shand and Alan Heirich.
Parallel Computing 29 (3), 311-323. (March 2003)
Abstract
Recently several distributed rendering systems have been developed
that exploit a cluster of commodity computers by connecting host
graphics cards over a fast network to form a compositing
pipeline. This paper introduces an algorithm for one such system to
parallelize contributions from arbitrary numbers of OpenGL lights, for
example to compute approximate soft shadows by sampling area light
sources. The algorithm renders multiple shadow maps across nodes of
the cluster by factoring the OpenGL lighting equation into
illumination and material properties. Additional parallelism is
achieved by utilizing multiple texture units within each graphics
card. Illumination by L point sources can be rendered by (L/K)+1 nodes
where K is the number of texture units on each graphics card. For
walkthrough applications each node requires a single rendering pass,
while for scenes with dynamic lights or geometry K+1 passes are needed
per node. We present results using fragment shaders on nVidia GeForce4
Ti4600 and ATI Radeon 8500 graphics cards, and a cycle-accurate
simulation of the compositing operators. These results show
interactive frame rates for walkthroughs and dynamic models rendering
shadows with 32 light sources on 9 nodes of a Sepia-2a distributed
rendering cluster.
Click here to order an
online copy of the paper.
Back to
Michael Isard's home page