Issues in the design of multi-server file systems to cope with load skew

Abstract

In a multi-server file system environment, related files (under the same subdirectory) are grouped together into filesets, and each filesets are assigned to a fixed server. The file operations to each fileset are served by its associated server. However, the dynamic load to various server varies with time and even with the best possible static load assignment, the loads to various servers are unbalanced (particularly in a small time interval). The short term sever load skew implies that extra server capacity has to be provided to satisfy response time requirement. In this paper, we make the observation that a large number of operations in many environments are to the files that are mostly-read and rarely updated. Such operations can be served by any server without incurring a high coherency. We also propose several dynamic load balancing algorithms of various degrees of complexity that achieve high reduction in load skew albeit reducing the server buffer hit probability. It is shown using real workload traces that server load skew can be significantly reduced using the above policies, and there is an inherent trade-off between load balancing and server buffer hit probability. Some variations of the proposed buddy based policies which limit the number of replications of buffer pages seem to provide a good compromise between the two factors.