(An) image parallel approach for ray tracing on distributed memory multicomputers

Abstract

Ray tracing technique has been widely used as the rendering algorithms for realistic image synthesis using global illumination model. It can simulate the global specular reflection but it has suffered from high computation complexity. Hence, many efforts for improvement performance have been presented through parallel implementation. Distributed memory multicomputer has good scalability in architecture and flexible programming model. But, its poor programming environment and relatively difficult programming technique are deficiencies to disturb the wider use. In recent years, it has started to be broadly used due to the improvement of programming environment and some commercial machines has been used in academic and industrial applications. Some researchers have studied on parallel ray tracing on distributed memory multicomputers. But, they show very different solutions according to the machines, and there remain unsolved problems yet. For parallel ray tracing, data distribution and runtime data transfer management scheme considering the secondary data structure should be studied and it should be decided which job distribution and load balancing scheme is efficient under the scheme. The objective of this thesis is to develop parallel ray tracing system and to evaluate on a real distributed memory multicomputer. We proposed a data distribution scheme to exploit the characteristics of the secondary data structure for accelerating ray object intersection. Hierarchical bounding volume tree is duplicated on all processors and primitives are distributed in interleaved fashion. Through the analysis and experimentations, we showed that our scheme could enhance the performance greatly without any overhead compared to other previous schemes. Additional memory requirement due to the duplication was offset by reducing cache area without performance degradation. Next, we experimented several job distribution and management schemes under our data distribution scheme. The res...