Performance Evaluation of a Two-Level Hierarchical Parallel Database System

Two typical architectures of parallel database systems are the shared-everything and shared-nothing architectures. Shared-everything architecture provides better performance than the shared-nothing architecture but it is not scalable to large system sizes. On the other hand, shared-nothing architecture provides good system scalability but is sensitive to data skew. Hierarchical architectures have been proposed to incorporate the best features of these two architectures (i.e., the shared-nothing and the shared-everything). In this paper, we present a detailed simulation study comparing the performance of the two-level hierarchical architecture with that of the shared-nothing and shared-everything architectures. The results from the simulation experiments presented here show that a properly designed hierarchical system can provide performance very close to that of the shared-everything while providing system scalability similar to that provided by the shared-nothing architecture. The results also indicate that, when designing a hierarchical system, the cluster size should be maintained as large as the hardware limitations would allow. Then, the performance of the hierarchical system would be very similar to that of the shared-everything architecture under the system and workload models considered here.