Parallel control technique and performance of an MPPM knowledge base machine

A description is given of parallel control techniques and performance evaluations by simulation for the knowledge-base machine (KBM) using the multiport page-memory (MPPM) and the unification engine (UE). Relational knowledge base retrieval requires repeated application of the unification-join (U-join) operations on term relations. It also requires each U-join to be dynamically scheduled, observing the result of the previously executed U-join. Control techniques to execute the query process of the relational knowledge base effectively in the KBM using the MPPM and the UE have been investigated. For parallel execution of coarse-grain operations, each operation is decomposed to concurrently executable fine-grain operations by partitioning input data sets. A number of features that depend on processor-allocation strategies have been identified. In particular, it was found that query processing resulting in repetitions of the retrieval operation, by a certain control strategy, increases the amount of the total data transfer to the highly parallel UEs and lowers processing efficiency. >