Performance analysis of the basic timestamp ordering algorithm via Markov modeling

Abstract

Locking and timestamping are two popular approaches to concurrency control in database systems. Although more than a dozen analytic performance studies of locking techniques have recently appeared in the literature, analytic performance study of times tamp-based concurrency control algorithms largely remains and unexplored area. Since a large class of concurrency control algorithms used timestamps, there is a strong need of analytic performance study of these algorithms. In this paper, we analyze the performance of the basic timestamp ordering concurrency algorithm for database systems. Due to the complex nature of concurrency control algorithms, it is very difficult to accurately analyze their perfomance mathematically. We show that even after making some assumptions, an exact performance model of the basic timestamp ordering concurrency control algorithm is so complicated that it is pratically impossible to find its closed-form solution. We circumvent this problem in the following way: We analyze a single transaction in isolation rather than analyzing the whole system. (This reduces the complexity of the analysis substantially.) We reflect the presence of other trasactions on the transaction under consideration by the probability that a conflicting access to a data object has been made. We assume that a transaction sees the average state of the system and all the transactions exhibit the average steady-state behavior, which permits us to work with averages rather than with probability distributions. Therefore, the technique used in the analysis is approximate. We illustrate the technique with numerical examples. We plot a set of performance measures for several values of the model parameters. The results of analysis are validated against a simulation study. We also extend the analysis to open database systems where transactions arrive from an external source.