Computation and Communication Aware task graph Scheduling on multi-GPU systems

Abstract

GPUs have emerged as popular throughput computing platforms due to the massively parallel computing capability and low cost. To attain further performance enhancement beyond single GPU, there is a growing interest in exploiting systems with multiple GPUs. Attaining superior performance in a multi-GPU system involves three main design challenges, namely load balance, memory utilization, and data transfer. Imbalanced loading across a system could cause idling of GPUs while poor data reuse would trigger excessive memory accesses. The inefficient data transfer between a host and a device becomes a considerable performance overhead during high throughput computing. This paper aims at addressing the above design issues by proposing a Computation and Communication Aware task graph Scheduling (CCAS) for multi-GPU systems. The proposed scheduling approach (CCAS) adopts an effective heuristic algorithm that considers both data reuse and load balance in a multi-GPU system. The data transfer overhead is hidden by extensively overlapping computation and data communication. The experimental results of the proposed CCAS have demonstrated an average of 22.15% performance enhancement when compared with a previous work.