Why a single parallelization trategy is not enough in knowledge bases

摘要

We address the problem of parallelizing the evaluation of logic programs in data intensive applications. We argue that the appropriate parallelization strategy for logic-program evaluation depends on the program being evaluated. Therefore, this paper is concerned with the issues of program classification and parallelization strategies. We propose several parallelization strategies based on the concept of data reduction—the original logic program is evaluated by several processors working in parallel, each using only a subset of the database. The strategies differ on the evaluation cost, the overhead of communication and synchronization among processors, and the programs to which they are applicable. In particular, we start our study with pure parallelization, i.e., parallelization without overhead. An interesting class structure of logic programs is demonstrated, when considering amenability to pure parallelization. The relationship to the NC complexity class is demonstrated. Then we propose strategies that do incur an overhead, but are optimal in a sense that will be precisely defined. This paper makes the initial steps towards a theory of parallel logic programming.