Conventional approachesto execution of database queries on general purpose multiprocessorsattempt to maximizesystemthroughputusinginter-queryparallelism with a fixed numberof processors.Standarduniprocessor optimization techniquesare usedto minimize execution time of individual queries, Our approachis to increase performanceby utilizing intra-queryparallelismaswell as minimizing overall resourcerequirements. Specifically, processor andtio bandwidth requirementsareminimized by coordinatingtheorderin which datapagesarereadinto memoryand pagejoins assignedto availableprocessors. Wepresenta schedulingstrategybasedonjoin indicesand provelower andupperboundson itsresourcerequirements. We thendescribea heuristicfor estimatingthenumberof processorsrequiredto completejoin executionin minimal time. Our simulation resultsindicatethatthesetechniques areeffectivewith respectto processorutilizationandbuffer requirements.
Marguerite C. Murphy, Doron Rotem