The Performance of Alternative Strategies for Dealing with Deadlocks in Database Management Systems

There is growing evidence that, for a fairly wide variety of database workloads and system configurations, locking is the concurrency control strategy of choice. With locking, of course, comes the possibility of deadlocks. Although the database literature is full of algorithms for dealing with deadlocks, very little in the way of practical performance information is available to a database system designer faced with the decision of choosing a good deadlock resolution strategy. This paper is an attempt to bridge this gap in our understanding of the behavior and performance of alternative deadlock resolution strategies. We employ a simulation model of a database environment to study the relative performance of several strategies based on deadlock detection, several strategies based on deadlock prevention, and a strategy based on timeouts. We show that the choice of the best deadlock resolution strategy depends upon the level of data contention, the resource utilization levels, and the types of transactions. We provide guidelines for selecting a deadlock resolution strategy for different operating regions.