利用LSF API实现GPU集群的并行机时统计  

A parallel computation time accounting method for GPU clusters using LSF API

作　　者：盛乐标 游伟倩[1] 张予倩 周庆林[1] Sheng Yuebiao;You Weiqian;Zhang Yuqian;Zhou Qinglin(High Performance Computing Center,Nanjing University,Nanjing,Jiangsu 210093,China)

机构地区：[1]南京大学高性能计算中心,江苏南京210093

出　　处：《计算机时代》2020年第11期63-65,69,共4页Computer Era

摘　　要：近年来GPU的使用越来越广泛,更多的高性能计算集群采用CPU和GPU的异构架构。准确的GPU计算机时统计,是大型计算机集群执行计费政策的基础,但作业调度软件LSF并未提供简便直观的GPU计算机时统计功能。文章提出利用LSF API结合C语言、MPI和Shell脚本语言等进行编程来解决,实现多样化、可定制的GPU计算机时统计和报表制作,为高性能计算中心和超级计算中心的管理带来便利。In recent years,GPUs have become more and more widely used,and more high-performance computing clusters have adopted heterogeneous architectures of CPU and GPU.Accurate GPU computation time accounting are the basis for large computer clusters to implement billing policies,but the job scheduling software LSF does not provide simple and intuitive methods for GPU computation time accounting.By using LSF API,programming with C language combined with MPI and Shell scripting language etc.is a good resolution,which can realize diverse and customizable GPU computation time statistics and make corresponding reports,so as to bring convenience to the management of high-performance computing centers and supercomputing centers.

关 键 词：高性能计算 集群管理 GPU集群 机时统计 

分 类 号：G482[文化科学—教育学]