Toward on-Line Predictive Models for Forecasting Workload in Clouds

Abstract

Forecasting workload plays a crucial role in term of Quality of Service (QoS) guarantee in cloud computing. However, there are existing challenges including continuous forecast and multi-step-ahead prediction due to the time series structure characteristics of workload. The most used approaches to forecast workload are predictive models based on time series analysis such as Auto-Regressive Integrated Moving Average (ARIMA) and Exponential Smoothing (ES) models, but these models do not address the limitations yet. This paper presents the strategies to apply machine learning to tackle the challenges of time series data streaming and develop on-line predictive models to continuously forecast and support the multi-step-ahead prediction. The empirical results show that the proposed approach yields much better accuracy than other methods, namely Multilayer Perceptron (MLP) and Long Short-Term Memory (LSTM).