On Disk Allocation of Intermediate Query Results in Parallel Database Systems

For complex queries in parallel database systems, substantial amounts of data must be redistributed between operators executed on different processing nodes. Frequently, such intermediate results cannot be held in main memory and must be stored on disk. To limit the ensuing performance penalty, a data allocation must be found that supports parallel I/O to the greatest possible extent. In this paper, we propose declustering even self-contained units of temporary data processed in a single operation (such as individual buckets of parallel hash joins) across multiple disks. Using a suitable analytical model, we find that the improvement of parallel I/O outweighs the penalty of increased fragmentation