Spatial Partitioning for Parallel Hierarchical Radiosity on Distributed Memory Architectures

Abstract

This paper presents an efficient, highly scalable implementation of the Hierarchical Radiosity Algorithm. We present a clever mapping of Hierarchical Radiosity to high-dimensional spaces that manifests a locality property, which can greatly reduce communication on parallel distributed memory architectures. We use a very simple dynamic spatial partitioning method to keep the mapping balanced. We describe solutions for the key implementation problems: asynchronous calculation, grouping of elements and links, and data reference locality. Speedup plots give an impression of the scalability of our implementation. On a Cray T3E the speedup curve is almost linear up to 64 processors. This is better than previously published attempts on massively parallel distributed memory computers.

Parallel Global Illumination Method Based on a Non-Uniform Partitioning of the Scene

“…The most recent works are being focused on distributed memory systems [15,10,6,7,11,1], either multicomputers or PC clusters. On such platforms the effective distribution of data among processors is very important.…”

“…In several proposals the scene is replicated on all processors [6,11], but this limits the applicability of the solutions to scenes with low complexity. A method valid for complex scenes implies the distribution of the scene among the processors [2,13,7,1].…”