Towards Efficient Parallel Radiosity for DSM-based Parallel Computers Using Virtual Interfaces
Résumé
This paper presents the performance evaluation of a new technique for radiosity computation which aims at exploiting efficiently the different levels of a memory hierarchy of both sequential and parallel computers. Such ability is essential when dealing with complex environments having several millions of polygons. The principle of our technique is to split the initial environment into several sub-environments and compute the radiosity within each sub-environment. Exchange of energy between sub-environments is performed by means of virtual interfaces and visibility masks. The size of sub-environments can be adapted in order to fit into a cache or a local memory. We performed several experiments using an SGI Origin 2000 to show the effectiveness of our solution. It improves both the sequential and parallel execution of a progressive radiosity algorithm. Our technique decreases the execution time on one processor of an SGI Origin 2000 by a factor of more than 5 and leads to a very good efficiency for complex environments (1 million of polygons) on a multiprocessor configuration.