Parallel Execution of Logic Programs

models for OR parallelism fall into three broad categories. The first, called “pure” OR parallelism, consists of a parallel search of the goal tree. The earliest system of this type was designed by Haridi and Ciepielewski [11]. When a searching process comes to a choice point, it can fork new processes for each alternative. Processes proceed from that point with very little interaction. If a process cannot unify its first literal with any program clause, it simply terminates; if it derives the null clause, it announces the result and then terminates. There is no need to send any information to another active search process. The second form of OR parallelism is based on objects called OR processes. In this model, a process corresponds to an object of Actors or Smalltalk [36, 41]. Each process is responsible for executing a small piece of the program. It maintains local state information, and communicates with other objects via messages. The third form, called search parallelism by Conery and Kibler [21], is based on a physical partitioning of the program. Clauses are stored in different nodes of a machine such as a multiprocessor database machine. The common aspect of the models in this category is that unification is done at the nodes where the clauses are stored, while the derivation of resolvents and overall control decisions are carried out by a different process. Two major issues in the exploitation of OR parallelism are the combinatorial explosion in the number of processes and representation of variable binding environments. Nondeterministic logic programs may have too much 38 Parallelism in Logic Programs