Design And Implmentation Of Generic Parallel Rendering System Based On GPU Cluster

Computer real-time photorealistic rendering technology has been widely applied to many fields, but when three-dimensional scene is very complex, a stand-alone computer is unable to meet the requirement for real-time applications. GPU-based parallel rendering technology becomes very important in computer graphics because of its capability to fully use the parallel computational capabilities of a GPU cluster. The parallel rendering framework based on GPU-cluster and dynamic load balancing strategy are studied in this paper. The following contents are included:1. The state-of-art of parallel rendering technology is introduced, which mainly focuses on parallel rendering framework and load balancing strategy.2. A generic parallel rendering framework based on GPU cluster is proposed. This framework used "RDTree" data structure to organize the division of tasks, composition of the rendering results, network communications, command dispatch. All the cluster machines could be divided into several different roles according to their responsibility, and each machine can dynamically transform from one role to another. The data transmission/reception, command execution, task distribution between machines is handled by CPU threads in parallel, and the rendering tasks are handled by GPU threads in parallel. The framework is easy to use and has good performance in versatility, scalability.3. A dynamic load balancing strategy based on crass-task is presented. This strategy is a past-frame strategy, and the rendering time of the past-frame is used to estimate the execution time of each task before the rendering of each frame, then the rendering resources allocated is calculated, and allocated for each task according to the proportion. This strategy can fully utilize the computational capabilities of the rendering resources, and achieve efficient performance.4. A GPU-cluster parallel rendering prototype system based on the above algorithm is designed and implemented. The experimental results show that our system has a good performance in system architecture and load balancing strategy.