Analysis of parallel processing architectures for database systems

Abstract

Two parallel architectures for relational database join operations are considered. Using a hierarchical decomposition technique, hybrid analytic/simulation models are developed to evaluate the effectiveness of the two architectures. To exploit parallelism, the join queries have to be partitioned into subtasks which can run on different processors in parallel. The impact of the stochastic nature of the task execution time on the amount of parallelism that can be exploited is studied. It is found that even if all the tasks have the same mean processing time, the variance can have an adverse effect on the performance improvement that can be achieved. Other factors, such as multiprogramming level, memory, communication bandwidth, etc., can lead to even lower speedup. This is addressed in the analysis. The study shows that it is more advantageous to couple a small number of large processors through a shared-memory architecture than to couple a larger number of small processors through a network.