Performance Prediction of a Parallel Monte Carlo Application: A Neural Network Approach

A feedforward neural network model is presented in this study to predict the execution time of a parallel Monte-Carlo implementation. The enormous performance range offered by today’s systems caused the performance evaluation tools to become more complicated to be able to consider the relative values and interrelated parameters. Artificial Neural Networks provide an excellent alternative to conventional techniques with their ability to capture many kinds of relationships and have been used successfully in various prediction tasks. However, their use in performance prediction area is a novel approach. The Neural Network model proposed here is aimed to be simple, general and reliable. This work also demonstrates the potential of artificial neural networks in identifying the contribution of interrelated system and application parameters to performance. Prediction of computational and communication execution times of the application is examined in this paper. Key-Words: Feedforward, Backpropagating, Neural Network, Performance Prediction, Parallel Systems, Monte Carlo