🤖 AI Summary
This paper addresses persistent high-amplitude oscillations in path-aware networks, arising from uncoordinated, greedy path selection by multiple endpoints. We quantitatively analyze their impact on efficiency, fairness, and convergence. To model the coupled dynamics of path selection and congestion control, we propose a novel dynamic system framework integrating game theory and control theory. We establish the first axiomatic analytical framework to formally classify periodic oscillation patterns and uncover a new mechanism enabling joint optimization of efficiency, convergence, and packet-loss avoidance. Theoretically, we show that user migration induces desynchronization, enhancing system stability—a finding that challenges conventional trade-off assumptions. Comprehensive simulations validate our theoretical predictions and yield quantifiable design principles for path-aware Internet protocols.
📝 Abstract
Path-aware networking, a cornerstone of next-generation architectures like SCION and Multipath QUIC, empowers end-hosts with fine-grained control over traffic forwarding. This capability, however, introduces a critical stability risk: uncoordinated, greedy path selection by a multitude of agents can induce persistent, high-amplitude network oscillations. While this phenomenon is well-known, its quantitative performance impact across key metrics has remained poorly understood. In this paper, we address this gap by developing the first axiomatic framework for analyzing the joint dynamics of path selection and congestion control. Our model enables the formal characterization of the system's dynamic equilibria-the stable, periodic patterns of oscillation-and provides a suite of axioms to rate their performance in terms of efficiency, loss avoidance, convergence, fairness, and responsiveness. Our analysis reveals a fundamental trade-off in protocol design between predictable performance (efficiency, convergence) and user-centric goals (fairness, responsiveness). We prove, however, that no such trade-off exists among efficiency, convergence, and loss avoidance, which can be simultaneously optimized through careful parameter tuning. Furthermore, we find that agent migration can, counter-intuitively, enhance stability by de-synchronizing traffic, a theoretical result validated by our simulations. These findings provide a principled design map for engineering robust, high-performance protocols for the future path-aware Internet.