An Adaptive Resource Provisioning Scheme for Effective QoS Maintenance in the IaaS Cloud

Effective bandwidth provisioning is of vital importance in the virtualized cloud where tenants with unique SLAs share a finite network. Different tenants collocated on the same physical server deployed with increasing VM density necessitates Quality of Service (QoS) provisioning beginning at the hypervisor. Recent efforts at provisioning the cloud network through various reservation methodologies have achieved some measure of success. However most of them do not account for the entire path over which application components communicate and cannot provide the necessary Service Level Agreement (SLA). Cloud applications components often communicate across multiple network devices aggregated into layers connected over finite bandwidth links that affect application response. Furthermore, traffic to and from tenant applications display volatility. In view of this, we design a virtual network reservation framework that is mindful of application performance across multiple network devices & traffic volatility. Our network reservation framework is based on a forecasting engine motivated by the volatility existent in traffic to and from virtualized cloud environments. This forecasting engine is able to maintain SLAs by employing dynamic time-series models to develop novel bandwidth provisioning thresholds that adapt to the time-variation in tenant workloads. We test the effectiveness of our methods in the OpenStack cloud environment focusing on traffic directionality in the datacenter network, VM density and QoS across multiple flows competing for finite bandwidth. Our forecasting method offers a 25% improvement in prediction accuracy over existing methods while the reservation framework maintains SLAs at 95%.