On the Inherently Speculative Nature of Nondeterministic And-Parallelism

proposed parallel execution model. This aspect concerns whether the model is to return: 1) one and only one answer to a query, regardless of how many exist, 2) any number of answers perhaps even all, but only one at a time, or 3) all answers to a query all at once. In Horn-clause logic programming, it is necessary for all subgoals in a clause to succeed for the clause itself to succeed. As a result, many models of and-parallel execution of logic clauses have generally considered it desirable to execute as many subgoals as possible in parallel. Contributing to this idea has been the belief that and-parallelism is not speculative, as is or-parallelism. This is, unfortunately, not true. This paper shows that "don't know" and-parallelism does exhibit speculative parallelism; it is in fact inherently speculative. Due to its speculative nature, executing subgoals in parallel can lead to superlinear speedups in a regular manner. Unfortunately, its speculative nature can also lead to serious slowdown rather than to speedup. Examples of scheduling activities which would be considered quite normal (even desirable) in non-speculative parallel systems are described and shown to be able to lead to slowdown. A "Slowdown Prevention Rule" is then described which, if adhered to, can ensure the absence of slowdown, or to at least bound its damage. Several additional aspects of speculative parallelism are also discussed. Models of the first type are "single solution" models; they exhibit "don't care" nondeterminism [Kowalski]; and they are typified by "committed choice", concurrent logic programming languages [Shapiro]. The second type are "any solution" (or, "many solutions") models; these models all exhibit "don't know nondeterminism" [Kowalski]. Parallel versions of sequential logic programming models, such as parallel Prologs, are included in this second type. Models of the third type are called "all solutions" models. These "all solutions" models have frequently been pursued for use in database query and/or implementation models, especially expert or knowledge data bases [Kerschberg]. This paper is concerned with the second type of parallel execution models — "any solution", "don't-know" parallel execution models. These models may exhibit andparallelism, or-parallelism, or both [Conery]. While both have proven useful, it has generally been felt that andparallelism was much better suited for highly deterministic programs. It has also been widely believed that, unlike orparallelism, and-parallelism in logic programs does not exhibit speculative computation (see [Tick], for example).