UC Santa Barbara

Programmers have come to expect better integration between databases and the programming languages they use. While this trend continues unabated, database concurrency scheduling has remained blind to the programs. We propose that the database client programs provide a large untapped information resource for increasing database throughput. Given this increase in expressive power between programmers and databases, we investigate how program analysis can increase database concurrency. In this paper, we demonstrate a predictive locking scheduler for object databases. In particular we examine the possibility to predict the client's use of locks. Given accurate predictions, we can increase concurrency through early lock release, perform deadlock detection and prevention, and determine whether locks should be granted before or during a transaction. Furthermore, we demonstrate our techniques on the 007 and other benchmarks.