A Parallel Simulation of Fully Connected Neural Network on a WK-Recursive Topology

The simulation of Artificial Neural Networks (ANNs) requires computing power that only massively parallel machines seem to be able to provide at reasonable cost. The choice of the interconnection topology is, certainly, one of the most important factor for the performances of the machines. In this paper, the order is to evaluate and compare a fixed-size WK-recursive topology for the distributed parallel simulation of fully connected neural networks. With the aim of obtaining performance improvements, in the design of the parallel algorithm, a systolic strategy has been persecuted, in order to guarantee the overlapping between elaboration and communication and the locality of communications, so that arouting is not required. After an introduction of the WK-recursive topology, performances evaluation has been undertaken and the results are compared with the ones obtained by using as target topology a ring of processors and a torus of processors. In particular, it is shown how WK-recursive configuration and the systolic strategy lead to performance enhancement for this class of neural networks.