Performance prediction of large MIMD systems for parallel neural network simulations

Abstract

Listen

Translate article

In this paper, we present a performance prediction model for indicating the performance range of MIMD parallel processor systems for neural network simulations. The model expresses the total execution time of a simulation as a function of the execution times of a small number of kernel functions, which have to be measured on only one processor and one physical communication link. The functions depend on the type of neural network, its geometry, decomposition and the connection structure of the MIMD machine. Using the model, the execution time, speedup, scalability and efficiency of large MIMD systems can be predicted. The model is validated quantitatively by applying it to two popular neural networks, backpropagation and the Kohonen self-organizing feature map, decomposed on a GCel-512, a 512 transputer system. Measurements are taken from network simulations decomposed via dataset and network decomposition techniques. Agreement of the model with the measurements is within 1–14%. Estimates are given for the performances that can be expected for the new T9000 transputer systems. The presented method can also be used for other application areas such as image processing.