Un lancer de rayons distribué pour traiter de très grands modèles sans précompilation

Abstract

This paper presents a distributed ray tracing algorithm in order to compute very large models (more than 100 millions triangles) with a cluster. The size of the used data set implies an out of core computation within a single computer. Ray tracing algorithm, present in many simulations of wave propagation, is an irregular process which imposes original solutions to manage efficient data distribution. Thus, the definition of a cache algorithm for main memory management improves data locality. Moreover, efficient data distribution on a cluster renders possible models whose memory needs are the sum of each local memory. This solution is freed from the traditional space preprocessing of the model and thus does not imply long latency before the visualization ofthe first rendered image.