机群管理系统中作业对象数据复制算法研究

A New Method for Replicating Scalable Data of Job Objects in Cluster Management System

机群管理系统中节点的I/O负载过重以及系统的可扩展性是制约其效率的关键。通过使用作业对象的互复制以及基于快照的并发调度,可降低读/写操作的等待,同时也使各种操作在机群节点交替执行,提高了并行性。文中首先描述了作业对象的快照模型;给出了作业数据访问协议并实现了作业对象可扩展复制算法;最后对该算法进行了评价和分析。

Aim. We propose a new data replication method based on a snapshot model in order to achieve the more efficient utilization of I/O resources. Section 1 in the full paper discusses job objects in cluster management system. Section 2 presents the replication protocol of job objects. Its two subsections are： the snapshot of job objects （subsection 2.1） and the protocol of job object operations （subsection 2.2）. In section 2, we reduce the waiting time for reading and writing through the concurrent scheduling based on snapshot. Section 3 explains the concurrent scheduling algorithm. In this section, we implement the scheduling algorithm to have concurrent access to job objects. The scheduling algorithm uses scheduler to check conflicts and only allows valid operations to be stored, thus ensuring the data consistency. In section 4, we did experiments on connecting 100 clients to the cluster server to track the three operations of job creation, job execution and job status inquiry. The experimental results, shown in Figs. 3 and 4, indicate preliminarily that the number of nodes of a replication cluster increases with increasing number of inquiries completed within a unit of time, thus consuming less response time and raising the efficiency by around 30%.