Main Memory DBMS on Modern Processors, a Simulation Approach for Database Performance Characterization

Database applications are an important type of workloads that is very different from other types of application workloads such as SPEC benchmarks. In the past, much research has been devoted to evaluate performance characteristics of database workloads on various processor architectures. While most of the previous performance evaluations have been done using hardware counters, many recently developed full-system simulators have provided another method for performance characterization. With the simulation paradigm, some architectural features of processors that were hard to examine using hardware counters are now accessible. However, results of simulations have to be validated against real machine executions to ensure their fidelity. In this paper, we extend the Simics full-system simulator to measure performance of architectural features of a simulated Pentium 4 alike PC running TPC-H workload on MonetDB (a main memory database management system). Then, we compare our simulation results with some well-established results published in literature previously. Our contributions are: (i) a flexible simulation methodology to characterize database performance; (ii) a close study of TPC-H benchmark performance including bandwidth measurement between memory and CPU, which has not been revealed in previously published research; (iii) validation of the simulation result against real machine measurement. Specifically, we find that our simulation results are mostly consistent with previously published results. In our simulation, approximately 83 % of query execution time in the database is spent on various stalls. Memory stalls and resource-related stalls are the most prominent ones for all TPC-H queries under test