Multivariate workload prediction using Vector Autoregressive and Stacked LSTM models

Abstract

Infrastructure as a service (IaaS) is a form of cloud computing services that provides virtual computing resources in the form of hardware, networking, and storage services to the end users as needed in an elastic manner. However, cloud-hosting platforms introduce several minutes delay in the hardware resource allocation. The obvious solution to this issue is to predict the future need of computing resources and allocate them before being requested. This paper represents a hybrid method for predicting multivariate workload based on the Vector Autoregressive (VAR) model and the Stacked Long Short Term Memory (LSTM) model. In the proposed method, two metrics are used: CPU and memory usage, the VAR model is used to filter the linear interdependencies among the multivariate time series, and the stacked LSTM model to capture nonlinear trends in the residuals computed from the VAR model. The proposed hybrid model is compared with other hybrid predictive models: the ARMLP model, the RNN-GRU model and the ARIMA-LSTM model. Results of experiments show superior efficacy of the proposed method over the other hybrid models.