Control of Communication Networks: Modeling, Simulation, and Optimization

本提案描述了一种设计网络控制算法的新方法，该算法结合了使用流量模型的在线仿真。该方法便于对仿真数据进行快速启发式分析，可应用于各种各样的网络控制问题，也可应用于各种各样的网络决策问题。pi有两个简单问题的概念验证实现：多类加权调度问题和为随机早期检测（RED）选择丢弃策略的问题。在这两种情况下，与以前的控制策略相比，它们的经验证明了实质上优越的性能。pi建议对其技术进行实质性的探索、分析和实现。他们计划将这些思想应用于网络决策问题，包括准入/访问控制、流量/拥塞控制、与代理服务相关的各种问题，以及诊断和恢复行动的选择。他们还计划将这些想法扩展到闭环系统，如TCP流，以及评估在利用各种流量模型（包括流体流量和远程依赖流量模型）时出现的问题。该计划利用了日益增长的流量建模工作，提供了使用此类模型影响网络控制性能的可能性。所提出的方法的广泛适用性为同时改进整个网络的各种控制算法提供了可能性。将这种控制方法与熟悉的模型推理技术一起使用，在不断变化的网络环境（包括攻击）下更新流量模型，也会自然地导致针对同样广泛的问题的自适应控制机制。

This proposal describes a novel approach for designing network control algorithms that incorporate online simulation using traffic models. The approach facilitates rapid heuristic analysis of simulation data, and can be applied to a wide variety of network control problems, and can be applied to a wide variety of network decision problems. The PIs have proof-of-concept implementations of this approach for two simple problems: the multiclass weighted scheduling problem, and the problem of selecting a dropping policy for random early detection (RED). In both cases, they have empirically demonstrated substantially superior performance when comparing to previous control policies. The PIs propose the substantial exploration, analysis, and implementation of their technique. They plan to apply these ideas to network decision problems including admission/access control, flow/congestion control, various problems relating to proxy services, and selection of diagnosis and recovery actions. They also plan to extend these ideas to closed-loop systems such as TCP flows, as well as to evaluate the issues that arise in utilizing various traffic models, including fluid-flow and long-range dependence traffic models. This plan exploits the growing work on traffic modeling, providing the potential to use such models to impact network control performance. The broad applicability of the proposed approach opens the possibility for simultaneous improvement of a wide array of control algorithms across the network. The use of this control approach together with familiar model inference techniques to update the traffic model under changing network environments (including attacks) also leads naturally to adaptive control mechanisms for the same wide variety of problems.