Practical Throughput Estimation for Parallel Databases

This paper describes an approach to performance estimation for shared-nothing parallel database systems. It estimates system throughput for a given benchmark or set of queries, and can exercise different data placement schemes to determine the data layout which provides the best throughput value. 1 Introduction Performance estimation is a valuable technique for the design and tuning of parallel database systems. It allows alternative design choices to be tested easily and cheaply and can highlight potential performance problems in designs long before any construction costs are incurred. A number of approaches have been developed to estimate performance using measures such as throughput or response time to handle a particular workload on some parallel database architectures. An example is APE (Adaptive Performance Evaluator) [1], which was designed to serve as the cost evaluation component for the query optimiser on DBS3 [2]. It includes four parts: KERNEL (a multi-environment cost evaluator of query execution plans), PROF (a statistical profiler of applications) , SPEC (a navigational user interface for environment specification) and PERF ( a graphical interface). APE uses an analytical model to estimate the response time for query execution. More accurate modelling of query execution can be achieved using simulation methods [3] but only at the cost of time-consuming simulation runs which do not directly predict average or typical behaviour