Operational Data Analysis: Improved Predictions Using Multi-computer Pattern Detection

Operational Data Analysis (ODA) automatically 1) monitors the performance of a computer through time, 2) stores such information in a data repository, 3) applies data-mining techniques, and 4) generates results. We describe a system implementing the four steps in ODA, focusing our attention on the data-mining step where our goal is to predict the value of a performance parameter (e.g., response time, cpu utilization, memory utilization) in the future. Our approach to the prediction problem extracts patterns from a database containing information from thousands of historical records and across computers. We show empirically how a multivariate linear regression model applied on all available records outperforms 1) a linear univariate model per machine, 2) a linear multivariate model per machine, and 3) a decision tree for regression across all machines. We conclude that global patterns relating characteristics across di erent computer models exist and can be extracted to improve the accuracy in predicting future performance behavior.