TimesLap: Mutability workload sequence prediction based on Laplacian Kernel in the cloud

Workload prediction is the key technology for elastic resource management in cloud platforms, and the prediction accuracy affects the efficiency of elastic management. However, a multitude of workload sequences have the characteristics of short-term mutation and nonlinearity in the cloud. A widely used strategy for predicting workloads is based on machine learning. The loss function of the existing workload prediction model cannot capture the nonlinear features well in the sequence. Moreover, it is sensitive to outliers and has low robustness, which affects the prediction accuracy. To address this issue, an integrated workload prediction method (TimesLap) is proposed based on Laplacian Kernel improved loss function(Laplacian Kernel MSE). Firstly, the workload sequence is decomposed into a high-frequency fluctuation sequence and a low-frequency fluctuation sequence. Then, the ARIMA model and LSTM-GRU model are applied to predict the high and low volatility sequences respectively. Laplacian Kernel MSE loss function is used to quantify the complex variability of high-frequency sequences, with the final prediction results are obtained by aggregating the prediction results of the models. Finally, the real trace of Google cloud and Microsoft Azure cloud are used for experiments. The experimental results show that TimesLap can effectively improve the generalization of model prediction. Compared with the state-of-the-art prediction methods based on Mean-Square Error, the Visualization Mean Absolute Error of prediction is reduced by 44%, and the \(R^{2}\) $R^2$ score is increased by 58%.