Transaction Support in Read Optimized and Write Optimized File Systems

Margo Seltzer Michael Stonebraker

Abstract

This paper provides a comparative analysis of five implementations of transaction support. The first of the methods is the traditional approach of implementing transaction processing within a data manager on top of a read optimized file system. The second also assumes a traditional file system but embeds transaction support inside the file system. The third model considers a traditional data manager on top of a write optimized file system. The last two models both embed transaction support inside a write optimized file system, each using a different logging mechanism. Our results show that in a transaction processing environment, a write optimized file system often yields better performance than one optimized for reads. In addition, we show that file system embedded transaction managers can perform as well as data managers when transaction throughput is limited by I/O bandwidth. Finally, even when the CPU is the critical resource, the difference in performance between a data manager and an embedded system is much smaller than previous work has shown.
Postscript Postscript