异构集群中CPU与GPU协同调度算法的设计与实现  

Design and implementation of CPU and GPU cooperative scheduling algorithm with heterogeneous clusters

作　　者：高原 顾文杰[1,2,3] 丁雨恒 彭晖[1,2,3] 陈泊宇 顾雯轩 GAO Yuan;GU Wen-jie;DING Yu-heng;PENG Hui;CHEN Bo-yu;GU Wen-xuan(NARI Research Institute,NARI Group Corporation(State Grid Electric Power Research Institute),Nanjing 211106,China;System R&D Center,NARI Technology Limited Company,Nanjing 211106,China;State Key Laboratory of Smart Grid Protection and Control,NARI Group Corporation,Nanjing 211106,China)

机构地区：[1]南瑞集团(国网电力科学院)有限公司南瑞研究院,江苏南京211106 [2]国电南瑞科技股份有限公司系统研发中心,江苏南京211106 [3]南瑞集团有限公司智能电网保护和运行控制国家重点实验室,江苏南京211106

出　　处：《计算机工程与设计》2020年第2期592-600,F0003,共10页Computer Engineering and Design

基　　金：国家重点研发计划基金项目（2017YFB0902600）;国家电网公司科技基金项目（SGJS0000DKJS1700840）大电网智能调度与安全预警关键技术研究及应用

摘　　要：为有效提高异构的CPU/GPU集群计算性能,提出一种支持异构集群的CPU与GPU协同计算的两级动态调度算法。根据各节点计算能力评测结果和任务请求动态分发数据,在节点内CPU和GPU之间动态调度任务,使用数据缓存和数据处理双队列机制,提高异构集群的传输和处理效率。该算法实现了集群各节点“能者多劳”,避免了单节点性能瓶颈造成的任务长尾现象。实验结果表明,该算法较传统MPI/GPU并行计算性能提高了11倍。To effectively improve the computing performance of heterogeneous CPU/GPU clusters,a two-level dynamic scheduling algorithm of CPU and GPU cooperative computing in heterogeneous cluster was proposed.Data were dynamically distributed according to each node’s computing capability evaluation result and task request,and tasks were dynamically scheduled between the CPU and GPU in the node,and data caching and data processing dual queue mechanism were used to improve the transmission and processing efficiency of heterogeneous clusters.The algorithm realizes that each node of the cluster follows the rule that the abler ones do more and avoids the long tail phenomenon of the task caused by the single node performance bottleneck.The experimental results show that the performance of parallel computing using this algorithm is 11 times better than that of traditional MPI/GPU parallel computing.

关 键 词：异构 集群 中央处理器 图形处理器 协同调度 算法 

分 类 号：TP305[自动化与计算机技术—计算机系统结构]