HierarchicalCDN: Federated Edge Intelligence with Metadata-Driven Cache Optimization for Live Streaming

Abstract

We present HierarchicalCDN, a novel CDN optimization system combining hierarchical LLM coordination with federated edge learning. The system employs three-tier orchestration (global, regional, edge) where edge LLMs operate autonomously while continuously learning from peer deployments across 50,000+ locations. Content-type specialization enables episodeaware prefetching for series (94% hit rate), event-window optimization for live sports (96% hit rate), and trailer-driven prediction for movies (88% hit rate). Federated learning enables edge nodes to share learned patterns without centralized retraining, achieving 23% improvement over isolated learning. Evaluation on production workloads serving 850 PoPs and 2.5 PB/day demonstrates 43% cache miss reduction, 37% latency improvement, 52% bandwidth savings, and 29% storage efficiency gains compared to traditional LRU caching.