Parallel inference performance prediction
Robert J. Bechtel, M.C. Rowe · IEEE Conference on Aerospace and Electronics · 2002
Following the work of Forgy and others on the RETE algorithm, minimal components of inference are characterized as having one or two inputs. Two-input components form inherent serialization points during inference, requiring both inputs to be available, and therefore forcing synchronization. Both one- and two-input components may have intercomponent dependencies which also impose sequentiality. Conflict resolution in the classical RETE model selects a single rule activation for execution from a possible large set of available activations, and therefore limits inference to a single pathway. An analysis tool which uses a grammar-driven parser to read rule files and construct a network of inference components is described. A dataflow analysis component then traverses this network, identifying points at which the dataflow imposes serial constraints. This analysis is provided to the user in report form, with a relative performance prediction for a specific parallel implementation.>