MDDS:一种面向高性能计算的并行文件系统元数据性能提升方法

发布时间：2018-01-20 15:02

本文关键词： 高性能计算并行文件系统元数据代理 I/O转发负载均衡　出处：《计算机研究与发展》2014年08期 　论文类型：期刊论文

【摘要】：随着计算能力的增强、应用课题规模和复杂度的增加,高性能计算机对并行文件系统性能要求越来越高.在海量小文件和大规模并发I/O操作的应用场景中,文件系统元数据的吞吐率成为限制其性能的关键因素.设计并实现了元数据代理(meta data delegation service,MDDS),通过降低元数据服务间的耦合度,保证元数据集群的高可用性;使用目录子树方式管理元数据代理空间,避免跨节点目录引入的分布式原子操作的复杂性和低效性.并针对高性能计算中I/O转发架构,提出基于元数据代理的两种作业调度策略——单作业独占单元数据代理调度和多作业共享多元数据代理调度——实现作业间和作业内的负载均衡.在116台存储服务器上对MDDS进行评估,实验结果表明,元数据代理提供了拟线性的元数据性能,在大规模的环境中较Lustre CMD方案有较好的扩展性;两种调度方式有效分散了作业元数据的负载,改善了高性能计算中的元数据瓶颈问题.
[Abstract]:With the enhancement of computing power and the increase of application scale and complexity, the performance of parallel file system is becoming more and more demanding by high-performance computers, in the application scenarios of large numbers of small files and large-scale concurrent I / O operations. File system metadata throughput has become a key factor limiting its performance. A metadata agent, meta data delegation service, is designed and implemented. MDDS can ensure the high availability of metadata cluster by reducing the coupling degree between metadata services. The metadata agent space is managed by directory subtree to avoid the complexity and inefficiency of the distributed atomic operation introduced by cross-node directories and the I / O forwarding architecture in high-performance computing. This paper proposes two job scheduling strategies based on metadata agent- single job exclusive unit data agent scheduling and multi job sharing multiple data agent scheduling to achieve load balancing between jobs and within jobs. The MDDS is evaluated on the storage server. The experimental results show that the metadata agent provides quasilinear metadata performance and is more scalable than the Lustre CMD scheme in large-scale environments. The two scheduling methods effectively disperse the workload of the job metadata and improve the bottleneck problem of metadata in high performance computing.
【作者单位】：江南计算技术研究所;
【基金】：国家“八六三”高技术研究发展计划基金项目(2011BAH04B03,2013AA013203)
【分类号】：TP38
【正文快照】： 高性能计算系统被广泛应用在科学研究、工业生产、国防军事等各种领域,在能源、航天、气象、生命科学、金融、动漫等各个行业发挥着重大作用,其按功能可分为存储子系统和计算子系统两部分.存储子系统为整机提供全局一致的共享存储空间,计算子系统以全局存储空间为纽带相互协作

【相似文献】