Disk Scheduling for Intermediate Results of Large Join Queries in Shared-Disk Parallel Database Systems

In shared-disk database systems, disk access has to be scheduled properly to avoid unnecessary contention between processors. The first part of this report studies the allocation of intermediate results of join queries (buckets) on disk and derives heuristics to determine the number of processing nodes and disks to employ. Using an analytical model, we show that declustering should be applied even for single buckets to ensure optimal performance. In the second part, we consider the order of reading the buckets and demonstrate the necessity of highly dynamic load balancing to prevent excessive disk contention, especially under skew conditions