Implementation and Performance of Cluster { Based FileReplication in Large - Scale Distributed

Large scale distributed le systems supporting 100s to 1000s of hosts are becoming increasingly important as distributed systems rapidly grow in scale. Our examination of the workload in one large commercial environment shows that widespread sharing of unstable les is also common in such large systems. This increase in scale and sharing places a heavy load on critical resources such as le servers and networks, leading to degradation in le access performance. Traditional client-based le caching techniques are inadequate in such environments. This paper describes the implementation and initial performance study of Frolic, a clusterbased dynamic le replication system. A cluster is a group of workstations and one or more le servers on a local area network. Instead of keeping copies of a widely shared le at each client workstation, these les are dynamically replicated onto the cluster le servers, so that they become locally available. We compare the performance of Frolic implemented on top of NFS with native NFS involving two clusters. File sizes and the number of consecutive accesses to a le from a particular cluster are varied in the sequential workload; workstation think times are varied in the concurrent workload. The results show that cluster-based le replication, as implemented in our Frolic prototype, can improve the performance of distributed le systems for a workload with wide-spread sharing of unstable les. 1