A comparison of transaction restart techniques in a distributed database environment

Abstract

The authors present the results of a simulation study carried out to compare the effects of different restart techniques on the overall throughput, number of restarts, average response time, and average communication delay in representative distributed database environments. The performances of the following transaction restart methods are compared: (1) restart with random increase of timestamp, (2) restart with random delay, (3) data-marking method, (4) data marking with random delay, and (5) restart with a substitute transaction. The substitute transaction method is shown to perform well under all loads, except for the case in which all transactions are update-only. In this case, restart with a random delay performs better. The data-marking method introduces a very high communication overhead owing to the fact that the transactions keep sending messages requesting operations when an item is not available. This results in high response times and in an erratic behavior of the system under higher loads.<>