A Dynamic Bandwidth Prediction and Provisioning Scheme in Cloud Networks

Effective resource provisioning for collocated cloud applications each with unique traffic patterns is a challenging task. Furthermore, resource requirements cannot be adequately determined a priori to deployment while traffic generated and received during application communication lifecycles is subject to fluctuations that exacerbate provisioning. Given this volatility, we analyzed traces from production cloud infrastructure mindful of the nonstationarity that presents challenges to Quality of Service (QoS) provisioning. Henceforth we developed a forecasting model that enables current predictive solutions widely applied in the cloud to adapt to a range of diverse operating conditions often observable as time-of-day load surges, traffic bursts and flash crowds resulting from sudden website popularity. We foster robustness in current methods by augmentation with a class of estimators adaptable to traffic volatility while providing a tradeoff between complexity and performance. The developed forecasting model avails itself of the Auto-Regressive Integrated Moving Average (ARIMA) model enhanced by a general class of Adaptive Conditional Score Models (ACS). The latter is able to adapt efficiently to load fluctuations observed variously as time-of-day traffic surges, flash-crowds and DoS episodes. The model offers between 10 & 15 % improved accuracy over existing models. We have realized a novel algorithm which has been adapted as a rate based solution applicable at integral points in the cloud network. The analysis and experimentation with our methods suggest a 25% reduction in Average Flow Completion Time (AFCT) when compared with current rate based methods of XCP & RCP while offering a 60% reduction in comparison to TCP.