In relational distributed databases a query cost consists of a local cost and a transmission cost. Query optimization is a combinatorial optimization problem. As the query size grows, the optimization methods based on exhaustive search become too expensive. We propose the following strategy for solving large distributed query optimization problems in relational database systems: (1) represent each query-processing schedule by a labeled directed graph; (2) reduce the number of different schedules by pruning away invalid or high-cost solutions; and (3) find a suboptimal schedule by combinatorial optimization. We investigate several combinatorial optimization techniques: random search, single start, multistart, simulated annealing, and a combination of random search and local simulated annealing. The utility of combinatorial optimization is demonstrated in the problem of finding the (sub)optimal semijoin schedule that fully reduces all relations of a tree query. The combination of random search and local simulated annealing was superior to other tested methods.

Two algorithms are developed for minimizing control overheads in exchanging state information arising from the control messages used in determining the load levels at other servers. In the first algorithm, the load levels at other servers are guessed using a simple heuristic algorithm. Such a model is found to provide significant improvements compared to the no-load sharing case. The second algorithm improves upon the first one by replacing some unnecessary task transfers by a single probe. The simulation results obtained from these algorithms are presented and compared to an algorithm based on random selection of destinations for transfer tasks. It was concluded that a load sharing policy should try to maximize the success rate in finding good destinations for transfer tasks while minimizing the control overheads.>