Parallel hash join algorithms for dynamic load balancing in a shared disks cluster

Most of previous parallel join algorithms assume a shared nothing (SN) cluster, where each database partition is owned by a single processing node. While SN cluster can interconnect a large number of nodes and support a geographically distributed environment, it may suffer from poor facility for load balancing and system availability compared to a shared disks sharing (SD) cluster. In this paper, we first propose a dynamic load balancing strategy by exploiting the characteristics of SD cluster. Then we parallelize conventional hash join algorithms using the dynamic load balancing strategy. We also explore the performance of parallel join algorithms using a simulation model of SD cluster. The experiment results show that the proposed parallel join algorithms can achieve higher potential for dynamic load balancing with the inherent flexibility of SD cluster.