NCDN: A Node-Failure Resilient CDN Solution with Reinforcement Learning Optimization

Abstract

Content Delivery Networks (CDNs) have enabled large-scale, reliable, and efficient content distribution over the Internet. Although CDNs have been very successful in serving a large portion of Internet traffic, they have several drawbacks. Despite their distributed nature, they rely on largely centralized management and replication. This can affect availability in case of node failure. Further, CDNs are complex infrastructures that span multiple layers of the networking stack. To address these issues, in this paper, we introduce NCDN, a novel highly distributed system for large-scale delivery of content and services. NCDN is designed to provide resilience against node failure through location-independent storage and replication of content. This is achieved through a two-layer architecture: the first layer (exposure layer) exposes services implemented by NCDN (e.g., Web, SFTP) to clients; the second layer (hidden layer) provides reliable distributed storage of content and application state. Content in NCDN’s hidden layer is stored and exchanged as Named Data Network (NDN) content packets. We employ the reinforcement learning (RL) to dynamically learn the optimal numbers of duplicates for different type of contents, because the RL agent has the advantage of not requiring expert labels or knowledge and instead the ability to learn directly from its own interaction with the world. The combination of NDN and RL brings NCDN fine-grained, fully decentralized content replication mechanisms. We compare the performance and resilience of NCDN to those of an idealized CDN via extensive simulations. Our results show that NCDN is able to provide higher availability than CDNs (between 8% and 100% higher under the same conditions), without substantially increasing content retrieval delay.