Join sequence optimization in parallel query plans

Abstract

We show an alternative way to optimize join sequences in a shared nothing architecture of the distributed relational DBMS HEAD. The hardware resources are also used for database independent applications in concurrency, for database access. To achieve a better hardware utilization we span a complete search space involving all intermediate shapes between left deep, right deep and bushy trees, but without join sequences including cartesian products. The approach is based on any combinatorial search engine and a complex cost function. The cost function involves the actual load of the processors, the network load, and the memory capacities. The new technique has free access to any join sequence in the whole search space. The approach is not restricted to acyclic join conditions.