云计算环境下多GPU资源调度机制研究

Study on Resource Scheduling Mechanism of Multi-GPU in Cloud Computing Environment

资源调度是云计算中的关键问题之一,它的调度机制与算法直接影响到云计算系统的性能及成本.GPU(graphics processing unit)正越来越多地被应用到通用计算领域,作为高性能云计算系统中的特殊计算资源,对GPU计算资源的调度有其特殊性.综合考虑计算任务在节点间以及节点内部的数据传输延迟,以充分利用系统GPU计算资源、掩藏传输延迟为目标,研究了云环境下多GPU的"传输&传输&执行"三段调度问题.提出一种云环境下GPU计算资源调度机制MGSC(Multi-GPU resource Scheduling scheme in Cloud environment):考虑了GPU计算中传输与计算的因素,讨论了在GPU计算中出现的四种资源需求情况,建立GPU计算资源模型;为了减轻中心节点的任务处理压力,设计了基于树型结构的GPU资源分布式检索算法.实验结果说明,MGSC在满足多用户共享GPU计算资源的同时,能够较好地提高云计算系统中GPU计算资源利用率,获得较高的服务质量,有效地减少资源闲置,降低服务提供者的服务成本.

Resource scheduling is one of the key issues in cloud computing.The performance and cost of cloud computing system are affected by scheduling mechanisms and algorithms directly.GPU（ graphics processing unit） is increasingly being applied to generalpurpose computing.As a special computing resource of high-performance cloud computing system,the scheduling of GPU computing resources has particularities.By considering the data transmission delay between computing nodes and inside the node,we focus on scheduling during the ＂Transport Transport Execute＂stages in multi-GPU cloud environment for the goal of taking full advantage of GPU and hiding the transmission delay.A GPU computing resource scheduling mechanism named M GSC（ M ulti-GPU resource Scheduling scheme in Cloud environment） is presented.The factors of GPU Data transmission and computing are considered.Besides,four cases in GPU computing are discussed and the computing resource model is established.In order to reduce the pressure of central node,a distributed GPU resource retrieval algorithm based on a tree structure is designed.Experimental results illustrate that M GSC can meet the requirements of multi-user in sharing GPU computing resources.In addition,M GSC improves the utilization of GPU computing resources to obtain a better quality of service and reduces the idle resources effectively and it reduces the costs of service provider.