An Adaptive Score Model for Effective Bandwidth Prediction and Provisioning in the Cloud Network

The collocation of applications in the cloud each with unique traffic characteristics presents provisioning challenges to providers. Furthermore, the volatility present in traffic sent & received by these applications necessitates a predictive solution for effective QoS strategies. To comprehend the tasks faced by cloud providers in this regard, we undertook statistical analysis of traffic traces measured periodically over a week, taking into account its nonstationarity. Henceforth, we developed a model to predict bandwidth utilization applicable in maintaining SLAs for multiple traffic flows at the cloud network edge and core. The developed univariate forecast model employs the Auto-Regressive Integrated Moving Average (ARIMA) model augmented with a general class of Adaptive Conditional Score Models (ACS). Our motivation for employing the ACS stems from its robust adaptation to outliers & transients more efficiently with increased computational accuracy than current methods; one of such methods being the recently adopted Generalized Auto-Regressive Conditional Heteroskedasticity (GARCH) to model volatility. These are observable typically as flash-crowds, time-of-day load spikes and DDoS events. The model offers improvements in forecasts over ARIMA/GARCH models with a 10-15% increase in predictive accuracy. Based on this model, we present 2 novel bandwidth adaptive algorithms which find application in the cloud network. The first is resident in the application layer of the SDN stack as a controller integrated into the OpenStack cloud environment. It was tested for effectiveness in a small test-bed where it sustains SLAs at 95%. The second methodology has been adapted in ns-2 where we present it as R²CP: Robust Rate Control Protocol, an extension to the Rate Control Protocol (RCP). R²CP offers a 25% reduction in flow completion time over XCP/RCP & over 60% reduction compared to TCP.