Scalable Prioritization of Restoration Activities in Highly Distributed CDNs with Cost and Performance Awareness
摘要
Highly distributed CDNs use their platform-level redundancy and failover mechanisms to reduce the uncertainty of the Return on Investment of nodes’ proactive maintenance. However, the Run-Till-Failure approach leads to hidden costs of using backups if the failure affects backup nodes’ IP transit 95/5 billing. Since CDN operations abstract from the topology of the underlying transport network, traditional graph-based, connectivity-focused network maintenance approaches are insufficient to address this problem. We propose a framework for prioritizing the restoration of caching nodes, minimizing the hidden cost while considering failure’s impact on delivery performance. Unlike traditional methods, performance is considered per-node. Impact assessment and minimization methods are included. The impact assessment method–Alternative-Cost-based Capacity Valuation Method (ALCOVE)–uses a novel, CDN-specific application of the Alternative Cost. Unlike in production environments, the Alternative Cost quantifies the monetary and performance difference between best delivery alternatives. ALCOVE enables drawing global conclusions from local analysis and seven orders-of-magnitude faster impact assessment than equally precise methods. We showed this mathematically and experimentally. The impact minimization method includes Integer Linear Programming (ILP) formulation of the prioritization problem and an exact algorithm, simplifying the computations–Monetary-Performance-Threshold (MPT). We solved the ILP for hundreds of failure scenarios in distinct CDN topologies based on the real-life Akamai platform. We compared ILP and MPT outputs with Pareto fronts of the most efficient and inefficient trade-offs between cost and performance impact minimization and related execution times. Results showed the exact nature of MPT, its scalability advantage over ILP, and a near-optimal trade-off: