摘要： 研究多集群的管理模式，设计支持多集群的用户决策和系统决策两种调度方式，提出基于PBS的多集群作业调度MCJSS架构。利用PBS的扩展接口，实现MCJSS的核心模块，管理多集群作业提交、作业转发和负载信息收集。给出了多集群间基于预测最轻负载转发的调度策略。运用两层负载信息收集策略服务于转发机制，利用组合单项负载的阈值来判断转发时机和目的集群。实验证明，相同的作业在相同规模但是不同组织形式的多集群和单集群组织模式下，多集群的系统吞吐量大于单集群组织模式。

关键词: 多集群, PBS系统, 组合负载值, 调度策略

Abstract: This paper studies the management model of multi-cluster, presumes two methods: user-decided and system-decided, and stabiles a MCJSS structure based on PBS. Implement the core of MCJSS to manage job submit, forward job and collect load information. It gives the schedule policy based on prediction of low load. Serve the job forwarding mechanism by using load information collection of two levels, judge the time to forwarding job and the destination. It is proved through experiment that the throughput of multi-cluster is higher than single model cluster model under submission of same number of jobs to same scale nodes.

Key words: multi-cluster, Portable Batch System (PBS), combination of workload, schedule policy

中图分类号: 

• TP391