An Isolation-Aware Online Virtual Network Embedding via Deep Reinforcement Learning

EasyChair Preprint 9871

Abstract

Virtualization technologies are the foundation of modern ICT infrastructure, enabling service providers to create dedicated virtual networks (VNs) that can support a wide range of smart city applications. These VNs continuously generate massive amounts of data, necessitating stringent reliability and security requirements. In virtualized network environments, however, multiple VNs may coexist on the same physical infrastructure and, if not properly isolated, may interfere with or provide unauthorized access to one another. The former causes performance degradation, while the latter compromises the security of VNs. Service assurance for infrastructure providers becomes significantly more complicated when a specific VN violates the isolation requirement.

In an effort to address the isolation issue, this paper proposes isolation during virtual network embedding (VNE), the procedure of allocating VNs onto physical infrastructure. We define a simple abstracted concept of isolation levels to capture the variations in isolation requirements and then formulate isolation-aware VNE as an optimization problem with resource and isolation constraints. A deep reinforcement learning (DRL)-based VNE algorithm ISO-DRL_VNE, is proposed that considers resource and isolation constraints and is compared to the existing three state-of-the-art algorithms: NodeRank, Global Resource Capacity (GRC), and Mote-Carlo Tree Search (MCTS). Evaluation results show that the ISO-DRL_VNE algorithm outperforms others in acceptance ratio, long-term average revenue, and long-term average revenue-to-cost ratio by 6%, 13%, and 15%.

Keyphrases: Internet of Things, Reinforcement Learning, Security, Smart City, Vertical Industries, isolation, resource allocation, virtual network embedding

@booklet{EasyChair:9871,
  author    = {Ali Gohar and Chunming Rong and Sanghwan Lee},
  title     = {An Isolation-Aware Online Virtual Network Embedding via Deep Reinforcement Learning},
  howpublished = {EasyChair Preprint 9871},
  year      = {EasyChair, 2023}}