Performance Study of Replica Concurrency Control Algorithms For Distributed Databases

Abstract

This study examines three replica concurrency control algorithms namely Distributed 2PL, Distributed OCC, and Distributed O2PL for distributed database systems. Four Different algorithms are performed are performed to evaluate the performance of above algorithms when they are incorporated with real-time data conflict resolution techniques namely PA, PB, PI, PA_PB. Among the four experiments, first experiment evaluates the performance of the various conflict resolution mechanisms (PA, PB, PI and PA_PB) when integrated with the 2PL and O2PL concurrency control protocols. Experiment 2 evaluates the performance of CC protocols based on the three different techniques: 2PL, O2PL and OCC. Experiment 3 is performed to evaluate the performance of these algorithms under different update frequencies. Experiment 4 is performed to evaluate the performance of these algorithms while varying number of replicas. Results of these experiments are analyzed and presented.

The performance metric employed for all experiments is MissPercent, the percentage of transactions that miss their deadlines. MissPercent values in the range of 0 to 30 percent are taken to represent system performance under “normal” loads, while MissPercent values in the range of 30 to 100 percent represent system performance under “heavy” loads. Several additional statistics are used to aid in the analysis of the experimental results, including the abort ratio, the message ratio, priority inversion ratio (PIR), and the wait ratio, which is the average number of waits per transaction. Further, the useful resource utilization is also measured as the resource utilization made by those transactions that are successfully completed before their deadlines.

All the missed deadline percentage for all experiments in this study is shown by graphs which only consider mean values that have relative half widths about the mean of less than 10% at the 90% confidence interval, with each experiment having been run until at least 10000 transactions are processed by the system.