The impact of workload on simulation results for distributed transaction processing

Abstract

To use benchmarks or to use real database traces (for the simulation of high performance distributed transaction processing), that is the question. Whether it is preferable to deal with special real workload, or to ignore the rich diversity observable with real measurements by restricting oneself to standardized benchmarks. We demonstrate how the evaluation and comparison of different load balancing algorithms depends on the workload considered. For that purpose we compare the TPC-B benchmark with two real workload traces: We analyze their properties with respect to workload characteristics on a theoretical level and we indicate how the differences observed influence performance evaluations in simulation experiments. The results imply that performance of different load distribution algorithms essentially depends on the workload processed.