Database concurrency control and recovery in local broadcast networks
W. Kevin Wilkinson · 1981
Some large business database systems are characterized by a high volume of short transactions (e.g. credit/debit account). In such systems, data retrieval costs are fixed and unavoidable. However, overhead due to concurrency control and recovery protocols may be reduced resulting in higher throughput and shorter response times. This research addresses the problems of concurrency control and recovery (collectively referred to as transaction management) in very large business applications. It is felt that traditional solutions to the problem (i.e. ever-larger centralized machines) will be inadequate as applications grow. An architecture for a database management system distributed over a local broadcast network is proposed. A concurrency control technique is presented which makes use of the broadcast nature of the communications bus. By eavesdropping for request messages on the broadcast bus, a single concurrency control node can perform conflict analysis for the entire system without explicit lock messages. Two algorithms are presented and shown robust with respect to communication and processor failures: a passive locking algorithm and a passive non-locking algorithm. Simulation results indicate that the passive schemes have very low overhead and perform better than corresponding distributed algorithms. Also, the cost of the recovery protocol necessary to ensure atomic commit at all sites (i.e. the distributed two-phase commit protocol) is shown to be quite high and, in many cases, overshadows the cost of concurrency control.