A Multi-step-ahead CPU Load Prediction Approach in Distributed System

Resource performance prediction is very important for resource management and scheduling in distributed systems. In this paper, we proposed a new multi-step-ahead prediction method for CPU load. It can be divided into three steps. The first step tries to find a function to fit the range change of the sequence. The second step is to predict the multi-step-ahead change (increase or decrease) pattern. We use multiple fixed length immediately preceding history sequences to obtain the change pattern prediction. Weighting strategies and machine learning algorithm are applied to synthesize different predictions that can be derived in terms of different immediately preceding history sequences with different lengths. Finally, change range prediction and change direction prediction are composed. Experiments showed our approach was more accurate than the approach of repeating one-step-ahead prediction to make the multi-step-ahead prediction, which is widely adopted in industry.