Task Scheduling for GPU Heterogeneous Cluster

Modern GPUs are gradually used by more and more cluster computing systems as the high performance computing units due to their outstanding computational power, whereas bringing node-level architectural heterogeneity to cluster. In this paper, based on MPI and CUDA programming model, we aim to investigate task scheduling for GPU heterogeneous cluster by taking into account the node-level heterogeneous characteristics. At first, based on our GPU heterogeneous cluster, we classify executing tasks to six major classifications according to their parallelism degrees, input data sizes, and processing workloads. Then, aiming to realize optimal mapping between tasks and computing resources, a task scheduling strategy is presented. The strategy consists of two key algorithms. The first is packing task algorithm (PTA) used to pack multiple tasks into a single task, such packing provides us a way of task classification converting according to the characteristic of computing resources. The second is system-level scheduling algorithm(SLSA) used to distribute parallel and sequential tasks to corresponding nodes, to maintain the load balance.