(The Eurographics Association, 2019) Dong, Yangzi; Peng, Chao; Childs, Hank and Frey, Steffen
Commodity workstations with multiple GPUs have been built by engineers and scientists for real-time rendering applications. As a result, a high display resolution can be achieved by connecting each GPU to a display monitor (resulting in a tiled large display). Using a multi-GPU workstation may not always produce a highly interactive rendering rate due to imbalanced rendering workloads among GPUs. In this work, we propose a parallel load balancing algorithm based on a screen partitioning strategy to dynamically balance the amount of vertices and triangles rendered by each GPU. Each GPU renders a screen region whose size may be different from the screen regions of other GPUs, but the amounts of vertices and triangles in those screen regions are balanced. It is possible that a screen region rendered by a GPU has to be displayed by another GPU. We propose a frame exchanging algorithm that allows GPUs to exchange screen regions efficiently. The inter-GPU communication overhead is very small since the data transferred between GPUs are a small amount of image pixels.