Parallel Join Algorithms on a Network of Workstations

We propose a non-traditional system architecture for backend database processing, i.e., a network of workstations. By moving data dynamically to otherwise idle workstations over a local area network, a high degree of processing parallelism could be achieved through load sharing. In this paper, three load sharing join algorithms are developed and implemented on a network of Sun workstations. These join algorithms are designed on the basis of different assumptions, e.g., whether indices are available and whether join result is required to be sorted. Based on the analyical results and empirical performance data for each algorithm, we are able to show that the load sharing strategy in each case is an effective method to achieve processing speedups. However, the network bandwidth of 10Mbits/second becomes a bottleneck, when more than a dozen or so workstations are deployed, which indicates that a higher bandwidth network is needed for further improvement in performance.