Performance evaluation of concurrency control algorithms for distributed database

Abstract

Among the different issues associated with a distributed database system (DDBS), perhaps the most difficult one is Concurrency Control (CC). Although a large number of CC algorithms are available in the literature, very few studies are concerned with the relative performance of these algorithms. Performance of CC algorithms is the central theme of this thesis. A locking algorithm and its variant are compared with a read?driven algorithm under a failure?free, fully redundant environment. A quantitative definition is proposed for the notion of the degree of conflict among transactions in a DDBS, and the utility of this notion as a performance index has been explored. The comparative study is based on simulations, using SIMPACK, on a DEC?1090 system.

A major conclusion of the study is that the read?driven algorithm performs better than the locking algorithms under the assumed conditions. A portion of the simulation results is subjected to regression analysis and optimization, and the utility of such an approach in database design is pointed out. Though the comparative study carried out here, like other major studies, assumes a failure?free environment, site and communication link failures are inevitable in a practical DDBS. The effect of such failures is to degrade the overall CC performance. With the objective of improving concurrency in a DDBS during failure periods, load?levelling based failure recovery techniques are developed. In doing this, a general distributed computing environment is considered, since a DDBS is only a special case of a distributed system. The conditions under which dynamic load?levelling is possible are derived.