Applying neural networks to computer system performance tuning

This paper presents results of empirical studies applying neural networks and techniques from control systems theory to computer system performance tuning. Experiments were performed on a simulated multiprogrammed computer system with a time-varying workload comprising four job classes. Key system performance measures such as device utilizations, mean queue lengths, and paging rates were collected and used to train neural network performance models. Several model-based adaptive control experiments show that backpropagation and radial basis function neural network controllers can be trained online to adjust memory allocations in order to meet desired performance objectives