On the partitionability of hierarchical radiosity

Abstract

The Hierarchical Radiosity Algorithm (HRA) is one of the most efficient sequential algorithms for physically based rendering. Unfortunately, it is hard to implement in parallel. There exist fairly efficient shared-memory implementations but things get worst in a distributed memory (DM) environment. In this paper we examine the structure of the HRA in a graph partitioning setting. Various measurements performed on the task access graph of the HRA indicate the existance of several bottlenecks in a potential DM implementation. We compare "optimal" partitioning results obtained by the partitioning software Metis with a trivial and a spatial partitioning algorithm, and show that the spatial partitioning copes with most of the bottlenecks well.

On the partitionability of hierarchical radiosity

“…For the one-room-scene and for other iterations / other processor number the situation is similar as in Fig. 4 [12].…”

“…In [12] we give data for the graph of the third iteration. This iteration is one of the computationally most complex iterations in terms of the number of processed links.…”

“…We start with the relative cut size for the nine room scene in the first and in the third iteration for varying processor numbers. Data for the one room scene is deposited in [12] because of space limitations. Figure 5 shows the results.…”

PHR: A Parallel Hierarchical Radiosity System with Dynamic Load Balancing

“…In a recent work [11], a parallel algorithm that uses spatial partitioning of patches to processors to improve locality and asynchronous calculation to hide latencies is presented. They report almost linear speed-up upto 64 processors on CrayT3E.…”

Spatial Partitioning for Parallel Hierarchical Radiosity on Distributed Memory Architectures

“…There have been quite successful attempts to implement HRA on a cache-coherent shared-address-space multicomputer [18], since HRA exhibits high fine grained parallelism. But developing a coarse grained parallel algorithm meeting the requirements of a distributed memory (DM) environment is difficult, which was confirmed by elaborate experiments in [11].…”