Rethinking traffic management: design of optimizable networks
Mung Chiang, Jennifer Rexford, Jiayue He · 2008
Traffic management refers to controlling how much traffic traverses each path in a network. On the Internet today, end hosts run congestion control to adapt sending rates, routers route traffic on shortest paths based on link weights, and operators tune link weights to direct traffic away from heavily-loaded links. This dissertation performs a top-down redesign of traffic management to support diverse application requirements, leveraging emerging technology trends in network virtualization and multipath routing. We begin by analyzing, then redesigning today’s traffic-management system. In the ‘bottom-up’ approach, we study the interaction of congestion control and traffic engineering using established optimization models. We find congestion control and traffic-engineering interact in a stable, though not always efficient manner. Efficiency can be improved by tuning the operator’s traffic-engineering function, but at the cost of robustness. In the ‘top-down’ approach, we propose a new objective function that captures the goals of both end users and network operators. Next, using various optimization decomposition techniques, we generate four distributed algorithms that divide traffic over multiple paths based on feedback from the network links. These distributed algorithms are provably stable and optimal, but can converge slowly and are sensitive to tuning parameters. Finally, combining the best features of these distributed algorithms, we construct TRUMP: TRaffic-management Using Multipath Protocol. TRUMP converges quickly and contains a single easy to tune parameter. Packet-level simulations show TRUMP behaves well with realistic topologies, feedback delays, capacities, and traffic loads. Since applications today may have different performance objectives, we next redesign traffic management to handle multiple traffic classes. A natural objective for an ISP is to maximize aggregate performance objectives across multiple traffic classes. Decomposing the ISP’s problem leads to a stable and optimal solution where each traffic class optimizes according to its own performance objective, with an algorithm to dynamically allocate bandwidth shares. The distributed protocols can be implemented using DaVinci: Dynamically Adaptive VIrtual Networks for a Customized Internet. In DaVinci, each virtual network runs traffic-management protocols optimized for a traffic class, and link bandwidth is dynamically allocated between virtual networks through separate queues. Overall, we show that using optimization theory as a foundation, simulations as a building block, and engineering intuition as a guide can be a principled approach to architecture and protocol design.