Efficient parallelizations of a competitive learning algorithm for text retrieval on the MasPar

In this paper, we present parallel implementations of a connectionist model for text retrieval on the MasPar MP-1, an SIMD machine with up to 16 K processors. The connectionist model was originally developed on a SUN SparcStation 1+ for a sequential implementation. In our parallel implementations, we consider three strategies for mapping the network onto the MasPar: one-to-one, many-to-one, and one-to-many, depending on the ratio of the network size to the number of processors, in order to reduce the computation time. We also consider load balancing among processors for further improvement in performance. Our experimental results demonstrate noticeable speedups in our parallel implementations.<
>