Mapping decisions by fuzzy inference

The approach presented is based on a system for mapping dynamic task tree-structures, such as occur in the relevant subfields of symbolic applications, to parallel machines. This mapping system provides multiple, and in many cases combinable, elementary strategies instead of a single universal one. The strategy configuration best matching the application characteristics, i.e. leading to optimal performance, can then be chosen. This requires establishing appropriate characteristics-oriented selection criteria that are expressive and precise enough to enable the compiler to find (close-to-)optimal configurations automatically. This paper focuses on the automatic-configuration aspect and presents the FiM system's solution to this task. FiM is implemented as a fuzzy-inference system, fuzziness allowing us to capture soft classifications of application characteristics and vague certainties or degrees of adequacy about the appropriateness of strategy selections. Existing approaches to fuzzy inference had to be extended to allow fuzzy multistage reasoning. The feasibility of the fuzzy-inference approach is shown. Though developed for mapping, the FiM approach can-using the corresponding selection rules-be applied to other configuration problems in multiple-strategy systems.