Cloud Server Load Turning Point Prediction Based on Feature Enhanced Multi-task LSTM

Abstract

Cloud workload turning point is either a local peak point which stands for workload pressure or a local valley point which stands for resource waste. The local trend on both sides of it will reverse. Predicting such kind of point is the premise to give warnings to the system managers to take precautious measures. Comparing with the value base workload predication approach, turning point prediction can provide information about the changing trend of future workload i.e. downtrend or uptrend. So more elaborate resource management schemes can be adopted for these rising and falling trends. This paper is the first study of deep learning based server workload turning point prediction in cloud environment. A well-designed deep learning based model named Feature Enhanced multi-task LSTM is introduced. Novel fluctuate features are proposed along with the multi-task and feature enhanced mechanisms. Experiments on the most famous Google cluster trace demonstrate the superiority of our model comparing with five state-of-the-art models.