| Research Abstract |
Recently, shared-nothing architecture is becoming the standard for commercial highly parallel machines such a SP-2, AP-1000, PARAGON etc. Thus, there have been a number of commercial parallel database systems implemented on these parallel machines or parallel environment. In order to exploit an inter-operation parallelism, there has been considerable effort focused on developing efficient multi-way join processing schedulers for parallel database systems. From the point of view of large database systems on a shared-nothing system these days, it is important for pipeline processing of multi-way join to transfer large amount of data through the network smoothly. So, parallel database processing on a shared-nothing architecture has to take the network bandwidth into account for query compilation, while most the researches so far done assumed shared everything architecture where network communication cost is neglected. Thus, it is necessary for multi-way joins in shared-nothing environment
… More
to exploit a new optimization algorithm considering the overlap of the I/O accesses and the network transfers.
In this research, we consider parallel multi-way join processing in a shared-nothing environment and show how multi-way join scheduling can be constructed efficiently for a given resource environment under finite network bandwidth. Although the communication bandwidth increases significantly these days, the advent of the disk array also increases the bandwidth of I/O data stream. This means the network could easily saturate when the number of join operation is not so small. Our algorithm first generates sub-tree seeds which fully consumes the network bandwidth in a pipeline processing. Then, these sub-tree seeds are combined each other to finally produce an optimal query tree. The restriction conditions for generating a short-tree which balances the I/O accesses and the network transfers and uses less memory are described in detail. The proposed algorithm for generating the multi-way join plan is evaluated in comparison with former work by using the introduced cost formula. From the evaluation results, not only is the quality of the proposed method better than previously presented algorithms such as left-deep, right-deep and segmented right-deep trees, but the quality of our algorithm does not deteriorate comparatively. Less
|
