Semi-supervised Neural Net Approach to Forecast Unpredictable CPU Load in an Enterprise Applications Environment

The aim of semi-supervised learning approach in this paper is to improve the supervised classifiers to investigate a model for forecasting unpredictable load on system and to predict CPU utilization in a big enterprise applications environment. This model forecasts the likelihood of a burst in web traffic to the IT systems and predicts the CPU utilization under stress conditions. The enterprise IT infrastructure consists of many enterprise applications running in a real time system. Load features are extracted while analyzing the patterns of work- load demand which are hidden in the transactional data of applications. This approach generates synthetic workload patterns, execute use-case scenarios in the test environment and use our model to predict the excessive utilization of the CPU behavior under peak load and stress conditions for the validation purpose. Expectation Maximization method with colearning, attempts to extract and analyze the parameters that maximize the likelihood of the model after subsiding the unknown labels. As a result of this approach, likelihood of excessive CPU utilization can be predicted in few hours as compared to few days. Workload profiling and prediction has enormous potential to optimize the usages of IT resources with low risk.