Proactive and Power Efficient Hybrid Virtual Network Embedding: An AWS Cloud Case Study

Abstract

The sharp increase of multimodal cloud resources demand makes it more challenging to design rightsized virtual instances. Inefficient embedding of high sized instances into the substrate resource network has led to numerous resource underutilization issues, which further constitute a key driver to repetitive reallocations of virtual instances. Besides, repetitive reconfigurations of virtual network instances go through a process of intra- or inter-cloud migration that provokes additional increase in power consumption. This paper proposes to solve these mutual challenges through a proactive, power efficient and hybrid Virtual Network Embedding (VNE) approach. We first formulated a Mixed Integer Linear Programming (MILP) model purposing to maximize total power efficiency of intra Data Center (DC) and inter networking resources as a function of EC2 instances requests rates. Leveraging the AWS cloud as a primary case study for this paper, the suggested VNE combines a multi-stage hybrid Virtual Node Embedding (VNoE) policy with an adaptive multistep consolidated Virtual Link Embedding (VLiE). As a starting point, a Green-Location aware - Global Topology Ranking (GLA-GTR) is designed as a primary ranking process suggesting the greenest substrate DCs locations with their related delivery networks. After implementing our proposal on a real AWS backbone network topology, simulation results indicated the efficiency of the proposed VNE approach. The Stacked Denoising Auto Encoders - Bidirectional Gated Recurrent Unit - Resources Vector Matching VNoE (SDAE-BiGRU-RVM VNoE) policy achieved a power decrease of 14.61%, 14.95% and 17.21% compared to BiGRU-RVM-VNoE, BiGRU-BF-VNoE and BF-VNoE policies, respectively. Accordingly, the suggested policy has reached significant power efficiency and overall maximized resource utilization.