SCORE: A Scalable Architecture for Implementing Resource Management Algorithms in High Speed Networks

支持逐流管理的能力——即维护状态、管理缓冲区、在每个流的基础上执行数据包分类和/或调度——提供了一个强大的工具来解决当今Internet中许多具有挑战性的资源管理问题，例如拥塞控制和端到端服务质量。然而，考虑到骨干路由器中的流量数量可以达到数十万个，每流管理机制所带来的复杂性可能会阻碍它们的有效实现和广泛部署。为了解决这些问题，首席研究员（PI）提出了一种称为可扩展核心（SCORE）的体系结构，在该体系结构中，网络被划分为称为岛屿的连续区域。在每个岛内，PI区分边缘和核心节点。边缘节点执行每个流管理，而核心节点不执行。由于边缘节点管理较少的流，并且在极限情况下，边缘节点可以是主机，因此该体系结构具有高度可扩展性。然后，主要目标是使用SCORE体系结构近似于一个参考岛，其中所有节点执行每个流管理。通过这种方式，可以为各种资源管理策略提供有效的支持。为了逼近参考网络，PI引入了两种方案。在第一种方案中，每个进入节点在每个数据包进入网络时都给它所属的流状态（例如，速率）打上标签。此外，每个核心节点根据其标签和节点状态（例如，利用率）处理数据包。在第二种方案中，将流路径上所有节点的状态聚合并发送到边缘节点。基于这些信息和流量状态，边缘路由器决定如何处理传入的数据包（例如，转发或丢弃它）。PI计划在SCORE框架内开发一套集成的算法，以支持最佳努力服务、端到端QoS保证和差分服务的拥塞控制。他将通过分析、模拟和实施相结合的方式进行研究。此外，他计划在CAIRN高速测试平台网络上进行大规模实验，并与工业合作伙伴（思科、Ascend、3Com、英特尔、MCI）合作，加快已开发技术向工业的过渡。这项工作的结果有望对互联网资源管理算法的设计产生重大影响。

The ability to support per-flow management -- that is, to maintain state, manage buffers, perform packet classification and/or scheduling on a per flow basis -- provides a powerful tool to address many of the challenging resource management problems in today's Internet, such as congestion control andend-to-end quality of service. However, given that the number of flows in a backbone routers can be on the order of hundred of thousands, the complexity incurred by per-flow management mechanisms may prevent them from being effectively implemented and widely deployed.To address these problems, the principal investigator (PI) proposes an architecture, called Scalable Core (SCORE for short) in which the network is partitioned in continuous regions, called islands. Within each island the PI distinguishes between edge and core nodes. While the edge nodes perform per flow management, the core nodes do not. Since edge nodes manage fewer flows and at the limit an edge node can be a host, this architecture is highly scalable. The main goal is then to use the SCORE architecture to approximate a reference island in which all nodes perform per flow management. In this way it is possible to provideefficient support for a wide variety of resource management policies.To approximate the reference network the PI introduces two schemes. In the first scheme, each ingress node labels every packet as it enters the network with the state of the flow (e.g., rate) to which it belongs. Further, each core node processes the packet based on its label and the node state (e.g., utilization). In the second scheme, the states of all nodes on the flow's path are aggregated and sent to the edge nodes. Based on this information and the flow state, the edge routers decide how to process an incoming packet (e.g., forward or drop it).The PI plans to develop an integrated set of algorithms within the SCORE framework to support congestion control for best-effort service, end-to-end QoS guarantee, and differential services. He will conduct the study via a combination of analysis, simulation, and implementation. In addition, he plans to conduct scaled experiments over the CAIRN high speed testbed network and work with industrial partners (Cisco, Ascend, 3Com, Intel, MCI) to speed-up the transition of the developed technologies to the industry. The results of this work are expected to have significant impacts on the design of resource management algorithms for the Internet.