Parallelization of a Backpropagation Neural Network on a Cluster Computer

M. Pethick, M. Liddle, P. Werstein, and Z. Huang (New Zealand)


backpropagation neural network, cluster computing, paral lelization, performance evaluation


This paper compares the performance of two paralleliza tion strategies for a backpropagation neural network on a cluster computer: exemplar parallel and node parallel strategies. Equations for calculating the theorectial costs of these two strategies are proposed based on the imple mentation presented in the paper. Performance results are collated according to different sizes of neural network, dif ferent dataset sizes, and number of processors. The perfor mance results show the advantages and disadvantages of the two strategies. More interestingly, we discover that the experimental results are very consistent with the theoretical costs. Therefore our cost equations can be used to predict which strategy is going to be better given a network size, a dataset size, and a number of processors.

Important Links:

Go Back