Adaptive Parallel Query Execution in DBS3

DBS3 (Database System for Shared Store) [1] is a parallel database system for sharedmemory multiprocessors [7]. It has been implemented on an Encore Multimax (10 processors) and on a Kendal Square Research KSR1 (72 processors). In a shared-memory architecture, each processor has uniform access to the entire database through a global main memory. Thus, the parallel scheduler, which allocates processors to the query's operations, has much freedom for balancing the query load onto processors. However, query response time can be hurt by several barriers which must be overcome by the scheduler: start-up time of parallel operations, interference and poor load balancing among the processors due to skewed data distribution. In DBS3, we use static relation partitioning to reduce interference and for compiletime query parallelization and dynamic allocation of processors to operations, independent of the degree of partitioning, to control start up time and load balancing. In this paper, we present the adaptive parallel query execution model of DBS3. A simple analysis outlines three important factors that influence the behavior of our model: skew factor, degree of parallelism and degree of partitioning. Then we reports on experiments with the 72-node KSR1 version of DBS3 with the relations cached in main memory. This main memory assumption is not a restriction of the model but a constraint of our KSR1 configuration which has a single disk.