On the performance of the immediate restart concurrency control policy

The authors study the performance of a concurrency control algorithm based on locking, called the immediate restart. Under this policy a transaction that fails in acquiring a lock is immediately absorbed (and replaced by another transaction to maintain the same multiprogramming level). The aborted transaction is attempted once again after some delay. The authors discuss the factors that influence the algorithm performance, namely data contention and resource contention. It is argued that selecting the appropriate delay is crucial to the performance of policy. A novel method for calculating the restart delay that takes into account the current state of system resources is proposed and shown to yield better performance than the traditional method which uses the running average of transaction response time under both uniform and nonuniform access