面向节点异构GPU集群的能量有效调度方案

摘要

GPU-enhanced clusters has become mainstream components in HPC field and are expected to be heterogeneous in node layer as the evolvement of processing elements (CPUs and GPUs) and the expansion of clusters nodes. In this paper, we propose an energy efficient task scheduling scheme for heterogeneous tasks in heterogeneous GPU-enhanced clusters. A formal description is presented to its task and resource model as well as energy consumption evaluation model in this paper. According to specific node selection policy, it can decrease the energy consumption loss of GPUs in idle status. By the division of task types and buddy allocation plus DVFS, it can improve the utilisation of CPU resource. What' s more, proceeding from system level, the scheme is compatible with algorithm-level and instruction-level energy optimisation.%GPU集群已经成为高性能计算(HPC)领域的主流组件.随着处理单元的发展和集群节点的拓展,GPU集群将在节点层面趋于异构化.提出一套针对异构任务在节点异构GPU集群上的能量有效调度方案.形式化地描述其任务和资源模型以及能耗评估模型.通过特定的节点选择策略,减少空闲状态的能耗损失.通过任务类型划分和组合分配以及DVFS,增加CPU资源利用率.该方案从系统层面着手,能够与现有的算法和指令层面的优化方法兼容.