Automatic Code Generation and Adaptive Grid Scheduling for GPU Cluster Computing

Recent advances in GPUs (graphics processing units) lead to massively parallel hardware that is easily programmable and widely applied in areas which require intensive computation besides graphics acceleration. The appearance of GPU clusters gains popularity in the scientific computing community, and the study on GPU clusters becomes an increasingly hot issue. While extending a singleGPU system to a multi-GPU cluster, the workload is spawned onto a number of GPU devices, and device-level and machine-level parallelism are achieved on top of the native SIMD thread-level parallelism. This requires a programming model extension and a workload scheduling strategy, but currently most of the programmers have to perform the extension manually and schedule the workloads in a naive way. Analytical performance models of GPUs exist, however, prediction of the performance of a GPU cluster is more complicated and no dedicated study has been done on this topic. In this paper, programming model is extended to fit the GPU cluster and propose a tool for automatic code generation and adaptive workload scheduling with which the application gains a 3x speedup from the cluster compared with executed on a single GPU. Besides, results show that with an extended analytical performance model, the performance of a GPU cluster is predictable.