A pattern fusion model for multi-step-ahead CPU load prediction

Abstract

In distributed systems, resource prediction is an important but difficult topic. In many cases, multiple prediction is needed rather than only performing prediction at a single future point in time. However, traditional approaches are not sufficient for multi-step-ahead prediction. We introduce a pattern fusion model to predict multi-step-ahead CPU loads. In this model, similar patterns are first extracted from the historical data via calculating Euclidean distance and fluctuation pattern distance between historical patterns and current sequence. For a given pattern length, multiple similar patterns of this length can often be found and each of them can produce a prediction. We also propose a pattern weight strategy to merge these prediction. Finally, a machine learning algorithm is used to combine the prediction results obtained from different length pattern sets dynamically. Empirical results on four real-world production servers show that this approach achieves higher accuracy on average than existing approaches for multi-step-ahead prediction.