Modern Datacenter Network Architectures: Spine-Leaf Topologies, Intent-Based Networking, and Intelligent Traffic Engineering

Abstract

The explosive growth of cloud computing, distributed artificial intelligence, and latency-sensitive applications has driven a fundamental transformation in datacenter network design, demanding architectures that deliver predictable low-latency, massive east-west bandwidth, and autonomous operational management at unprecedented scale. This paper presents a comprehensive investigation of modern datacenter network architectures across three interconnected domains: spine-leaf topologies and their evolution beyond classical Clos designs, intent-based networking (IBN) as a paradigm for declarative network management, and intelligent traffic engineering powered by machine learning and reinforcement learning. We introduce the Smart Intent-driven Traffic Engineering (SITE) framework, an integrated architecture that unifies topology-aware forwarding, intent compilation and verification, and deep reinforcement learning-based traffic optimization into a closed-loop autonomous networking platform. Evaluation across simulated datacenter fabrics spanning 2,048 to 32,768 endpoints demonstrates that SITE achieves 93.7% of theoretical optimal throughput under heavy load (compared to 58.4% for ECMP), reduces P99 tail latency by 52.3%, lowers network-induced application latency by 38.7%, and enables fully automated intent-driven network changes with 99.2% correctness. We further analyze the architectural trade-offs between topology variants (spine-leaf, fat-tree, dragonfly, rail-optimized), the maturity levels of intent-based systems, and the effectiveness of machine learning techniques for traffic prediction, congestion avoidance, and load balancing. Through systematic evaluation against 61 references spanning academic research, industry standards, and hyperscaler deployments, this article provides network architects and researchers with a unified reference for designing, operating, and evolving modern datacenter network infrastructures.