On coupling multiple systems with a global buffer

Abstract

We conduct a performance study of coupling multiple systems with a global buffer, and present several results obtained from a multiple-system simulator. This simulator has been run against three workloads, and the coupled system behavior with these three different inputs is studied. Several statistics, including those on local and global buffer hits, page writes to the global buffer, cross-invalidations, and castouts are reported. Their relationship to the degree of data skew is explored. Moreover, in addition to the update-caching approach, a design alternative for the use of a global buffer, namely read-caching, is explored. In read-caching, not only updated pages but also pages read by each node are kept in the global buffer, thereby facilitating other nodes access to the same pages at the cost of a higher global buffer usage. Also investigated is the case of no-caching, i.e., without using a global buffer. Several simulation results are presented and analyzed.