ITR: Multiple Time Scale Traffic Control for Next Generation Internets

互联网由一系列协议组成，这些协议包括拥塞控制、路由、组播、标签交换、资源预留和准入控制、差错控制和网络管理。这些协议在多个时间尺度上执行控制操作，从用于路由和标签交换的微秒，到用于拥塞控制的毫秒，以及用于组播、路由表更新和网络管理的几秒范围。网络系统的行为和性能受到驱动协议的工作负载的影响--重尾文件的可靠传输、VBR视频和音频的对服务质量敏感的流传输、连接到达的突发性、短流和长流的倾斜组合以及多路传输流量的自相似突发性--互联网工作负载在多个时间尺度上表现出可变性和相关性，从观察到的毫秒范围内的IP流的乘性扩展，到第二范围内的长范围相关性。随着服务质量成为统一的保护伞，各种网络子系统可以被组织起来并被视为提供用户指定的服务，因此出现了协调和集成控制活动的需求，其中许多控制活动直接影响端到端的服务质量，从而既可以提供有效的服务，也可以提供高效的服务。伴随着服务质量集成挑战，PI有机会明确地利用网络协议和互联网工作负载的多时间尺度特性来影响看不见和可预测的服务。在拟议的项目中，PI计划解决以下双管齐下的问题：(1)利用网络协议的多时间尺度特性来促进不连续的网络控制的有效协调和集成，以实现端到端的服务质量；(2)利用互联网工作负载的多时间尺度特性来实现对工作负载敏感的流量控制。问题(1)又由两个关键问题组成：充分性或分离条件，在这种条件下，作用于不同时间尺度的两个网络控制--例如，路由和拥塞控制--可以在不对稳定性和效率造成有害影响的情况下被集成，以及当时间尺度分离不可用时，协议的有效耦合--例如，标签控制和拥塞控制。问题(2)包括三个关键问题：使用轻量级乐观控制的短连接管理，使用连接时长预测和多层反馈控制的长连接管理，以及通过对工作负载敏感的端到端和逐跳控制来提高服务质量。PI计划攻击的双重挑战基于明确定义的技术问题，同时代表了创新的流量控制维度，对下一代互联网具有更广泛的影响。PI带来了核心流量控制领域的专业知识，包括拥塞控制(Park)、Qos Routing(HOU)、多播(HOU)、标签交换(Park)和流量建模(Park)。这两个PI在性能分析和协议设计方面都有丰富的经验，在基于仿真(HOU)和实现(PARK)的性能评估方面互补优势。PI将利用现有的基准测试平台--由Cisco 7206路由器(PARK)和NetSim性能评估环境(HOU)组成的普渡Infobahn服务质量测试平台--来实施、测试和基准测试协议。研究成果、技术演示和软件原型将通过Web提供给学术机构、行业分支机构和整个网络社区。

The Internet is governed by a compendium of protocols spanning a number of functionalities including congestion control, routing, multicast, label switching, resource reservation and admission control, error control, and network management. These protocols engage in control actions at multiple time scales, from microseconds for routing and label switching, to milliseconds for congestion control, and the several second range for multicast, routing table updates, and network management. The behavior and performance of a network system is influenced by the workload that drive the protocols - reliable transport of heavy-tailed files, QoS-sensitive streaming of VBR video and audio, burstiness of connection arrivals, skewed make-up of short- and long-lived flows, and self-similar burstiness of multiplexed traffic - with Internet workload exhibiting variability and correlation at multiple time scales, from multiplicative scaling observed for IP flows in the millisecond range, to long-range correlations present in the second range. With QoS emerging as a unifying umbrella under which the various network subsystems canbe organized and viewed as provisioning user-specified services, the need arises to coordinate andintegrate the control activities - many of whom have direct impact on end-to-end QoS - such thatboth effective and efficient services can be rendered. In tandem with the QoS integration challenge, the PIs are presented with the opportunity to explicitly exploit the multiple time scale nature of network protocols and Internet workload to affect seemless and predictable services. In the proposed project, the PIs plan to address the following two-pronged problem: (1) exploit multiple time scale property of network protocols to facilitate effective coordination and integration of disjoint network controls for end-to-end QoS, and (2) exploit multiple time scale nature of Internet workload to achieve workload-sensitive traffic controls. Problem (1), in turn, is comprised of two key issues: sufficiency or separation conditions under which two network controls - e.g., routing and congestion control-acting at different time scales can be integrated without causing harmful effects with respect to stability and efficiency, and the effective coupling of protocols when time scale separation is not available - e.g., label control and congestion control. Problem (2) consists of three key issues: short-lived connection management using lightweight optimistic control, long-lived connection management using connection duration prediction and multi-layered feedback control, and QoS amplification through workload-sensitive, end-to-end and per-hop control. The two-fold challenge that the PIs plan to attack is grounded in well-defined technical problems, and at the same time, represents an innovative traffic control dimension with broader implications to the next generation Internet. The PIs bring expertise in core traffic control areas spanning congestion control (Park), QoSrouting (Hou), multicast (Hou), label switching (Park), and traffic modeling (Park). Both PIshave significant experience in performance analysis and protocol design, and they complement each other's strengths with respect to simulation (Hou) and implementation (Park) based performance evaluation. The PIs will leverage existing benchmarking platforms - the Purdue Infobahn QoS testbed comprised of Cisco 7206 routers (Park) and the NetSim performance evaluation environment (Hou) - to implement, test, and benchmark the protocols. The research results, technology demonstration, and software prototypes will be made available through the Web to academic institutions, industrial affiliates, and the networking community at large.