Abstract—Forecasting resource usage values of a cloud service has ample applications such as service performance management, auto-scaling, capacity planning, and so on. While univariate forecasting techniques are the focus of current research, multivariate forecasting is rarely explored. This research work focuses on multivariate forecasting of resource usage values believing that there exists interdependency among the features of the underlying system that must be considered while forecasting. At first, the interdependency among the attributes is verified using Granger causality tests. Then the research explores various forecasting approaches — univariate Multi-Layer Perceptron (MLP), univariate Long Short Term Memory (LSTM), multivariate Vector Autoregression (VAR), and multivariate stacked LSTM. Further based on the observations of performances of these models the research proposes an implementation of a weighted ensemble of VAR and LSTM models to forecast key cloud resource usage metrics. The models thus proposed are implemented and validated using the publicly available GWA-T-12 Bitbrains time series dataset. The results show that the multivariate models outperform univariate models with lesser Normalised Root Mean Square Error (NRMSE) values. Also, the multivariate stacked LSTM outperforms VAR and the proposed ensemble forecasting model with lesser NRMSE values within a range of 1–5% for various resources across different lag values.
 
Keywords—multivariate forecasting, cloud resource usage forecasting, Long Short Term Memory (LSTM), stacked LSTM, Vector Autoregression (VAR), ensemble forecasting


Copyright © 2023 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC BY-NC-ND 4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.