File Migration in Distributed Computer Systems

摘要

The management of global (shared) files can have a big impact on the performance of distributed computer systems. In this dissertation effort is devoted to the study of the file referencing process and to the development of algorithms to place and migrate files in distributed computer systems. The topic of assigning files to the nodes of computer networks has received considerable attention in the past. Unfortunately, most of the work in this area has concentrated on the optimization of distributed file systems based on extremely simple models of the workload. This study is based on traces of activity collected from a number of real systems. The workload could be analyzed and algorithms adapted to it. Synthetic distributed systems were obtained from the real centralized systems that generated the traces. The SLAC and the Hughes systems were used. This partitioning of the users of a large centralized computer system into a number of smaller user communities was based on their utilization of shared files. The SLAC system was found to produce user subsets with a higher degree of overlapping (in terms of shared files) than the Hughes Aircraft installation. The file migration policies are introduced. First the single copy policies are presented. In this context a migration policy was developed that is optimal in the sense of minimizing the average network traffic. The existence of extended localities in the use of shared files suggested the use of multiple-copy policies in order to reduce the network traffic even further. A number of policies based on the Working Set model were studied. The best performing policies in terms of generated traffic are those that closely monitor the update traffic and that react to it by deleting the copies that are too expensive to maintain up to date. (ERA citation 07:060691)