Neural Network based Performance Prediction with Feature Extraction

In distributed systems, efficient utilization of resources is a big challenge. The status of resources continuously changes and is hard to keep track of. A predictive approach can solve this problem by forecasting resources' status based on their historical performances. In this paper, such an enhancement is analyzed: the utilization of resources is periodically monitored and future utilizations are predicted based on this historical information. These predictions can be used by the scheduler when making job assignment decisions with specific resource requirements. They can also be used for detecting future bottlenecks and failures in the network. In this paper, a feature extraction and neural network combined approach is analyzed: features are extracted for efficiency and faster results. Static, linear, nonlinear, dynamic and recurrent networks are analyzed for time series prediction of resource's performances. Recurrent networks combined with wavelet feature extraction process resulted best predictions