Performance analysis of locking methods with limited wait depth

Abstract

A number of recent studies have proposed lock conflict resolution methods to improve the performance of standard locking, i.e., the two-phase locking concurrency control method with the general waiting policy. The wait depth limited (WDL) method, which is an improvement on the running priority (RPS) method, outperforms other known concurrency control methods, except optimistic concurrency control methods in a system with "infinite" processing capacity. This paper provides an analysis of RPS, which is extensible to modified WDL with a performance comparable to WDL, but requiring fewer comparisons to determine the transaction to be aborted. The analyses of RPS and modified WDL are much more complicated than those of the no waiting and the cautious waiting methods, which only abort transactions making conflicting lock requests, while both active and blocked transactions can be aborted by RPS and modified WDL. We also extend the analysis to a system with multiple transaction classes, e.g., as determined by the number of requested locks, which occur according to prespecified frequencies, i.e., the fractions of transactions in the input stream. Our analysis ensures that the composition of completed transactions as affected by transaction aborts is not different from the incoming transaction frequencies. It follows from validation results that the analysis of RPS is quite accurate up to the point at which the peak throughput is attained in an infinite resource system. We also demonstrate the correspondence between flow diagrams and semi-Markov chains and provide justifications for using the exponential distribution for transaction waiting times. © 1998 Elsevier Science B.V. All rights reserved.