DLVisor: Dynamic Learning Hypervisor for Software Defined Network

Abstract

Software Defined Network (SDN) is one of the modern networking technologies that provide network flexibility and simplifies network management. Virtual SDN (vSDN) enhances the flexibility of sharing physical networking resources by multiple slices representing multiple tenants or services where each tenant has control over their services or applications over the Virtual Network (VN). Network virtualization gives service providers more flexibility to offer new and innovative services with extra efficiency and reliability. Running multiple virtual networks over a given infrastructure creates challenges for efficient resource allocation mechanisms to avoid congestion and resource starvation and to maintain Service Level Agreement (SLA), where resource management in vSDN is carried out by hypervisors. Few studies have addressed dynamic resource allocation in the vSDN domain. Therefore, to efficiently utilize the resources of the virtualized networking infrastructure, network hypervisors must be proactive with self-reconfiguration capabilities to assign the physical resources and be highly adaptable and react to changing vSDN future demands. Thus, dynamic learning-based hypervisors are to improve hypervisor operations. Based on that, this study aims to enhance the vSDN technology to provide an enhanced proactive dynamic slice resource allocation mechanism, to improve traffic delivery and resource utilization. This can be fulfilled by proposing an enhanced intelligent forecasting model for vSDN slice resource utilization based on improved statistical and Machine Learning (ML) techniques. The proposed model will react dynamically to the concept drifts and then be utilized to develop a resource allocation mechanism for vSDN slice resource allocation. The improved dynamic forecasting resource allocation mechanism is verified through available real network traces datasets from various sources. The DLVisor with its Dynamic Learning Framework (DLF) can reduce overutilization and, consequently, resource starvation by 100% compared to the related benchmark.

Keywords

• Machine learning
• resource allocation
• resource forecast
• software defined network
• virtualization