Distributing backward-chaining deductions to multiple processors

Abstract : It is widely believed that parallel computation will be the basis for the next major advance in computing speed. In reality, many difficult problems remain to be solved. Two such problems are addressed in this thesis: (1) the design of a parallel execution model that exploits desirable types of parallelism; and (2) the design of a resource allocator to map the parallel computation to hardware resources for processing, storage, and communication. The thesis presents a parallel execution model called PM for backward-chaining deduction with Horn clauses. For the target multiprocessor class, PM can exploit the most parallelism among existing execution models that use data-driven control. In particular, PM can exploit or-parallelism, and-parallelism, and pipelining. The target class of multiprocessors has the following properties: (1) there are an arbitrary number of MIMD processors; (2) each processor has some local memory but there is no global memory; (3) processors can communicate only by sending messages to each other; (4) message delay is a function of the amount of data in the message and the distance between source and destination; and (5) each processor can perform backward-chaining deductions based on the subset of the program that it contains. The type of backward-chaining deduction is restricted. In particular, no recursive clauses are allowed, unit clauses must be ground, and certain probabilistic uniformity and independence assumptions must apply. Second, a partitioning of the database is assumed to be given. Both greedy allocation and local minimization are based on a formally defined cost function that quantifies intuitive notions of undesirable allocations. Algorithms are presented for the efficient computation and recomputation of this cost function. Considerable speedups are obtained by using this allocation strategy.