Fine-Grained Traffic Management in Computer Networks

This dissertation investigates fine-grained traffic management in computer networks. Traffic management can effectively help networks to achieve higher Internet reliability, efficiency, and performance. As today't Internet traffic is growing rapidly and more dynamically, traffic management technologies that perform in fine grained become increasingly appealing because they manage and control network resources dynamically and adapt to the real-time traffic and condition changes. In this dissertation, we investigate fine-grained traffic management in three different dimensions: time, space, and application. In the time dimension, we explore high-frequency traffic engineering. There exists an intrinsic performance tradeoff between responsiveness and stability for adaptive traffic engineering that performs in high frequency. We analyze it from a feedback control perspective, and derive a model that characterizes the system parameters' effects on the performance of the dynamic routing system. This allows quantitative analysis of adaptive TE algorithms and their design parameter choices. We then specialize the general framework in two representative network topologies and derive the stability conditions for their dynamic routing systems. Together they provide systematic insights on the relations among several network factors and the intrinsic tradeoff among different network control objectives. In the space dimension, we investigate fine-granularity traffic split. For given traffic split ratios calculated mathematically by routing algorithms in the routing engine, the routing realization mechanisms in the data plane implement such splits without breaking flows. Treating all flows equally, the state-of-the-art approaches deployed in switches do not provide enough accuracy especially when facing non-uniform flow size distribution. To accurately realize given traffic split ratios in switches with small performance degradation, we instead propose a dynamic load distribution scheme based on the collected load sharing statistics and incorporate such tradeoff in a weighted sum optimization. It finds the most accurate traffic splits with minimum route changes. In the application dimension, we focus on per-application end-to-end path selection. The standard way to obtain end-to-end SLAs by creating private networks through business contracts is costly and takes lots of time to realize. We propose a platform that selects end-to-end path based on application specific performance need in real time through overlay networks. With the knowledge the network topology and conditions, it strives to achieve the optimal end-to-end performance by exploring the last-mile diversity. It allows the flexible and responsive per-application or per-end-user selection of the edge node for the overlay networks, and thus can fast recover from network failures and performance degradation. We present our design of the end-to-end throughput optimization system with detailed discussion of each component including dynamic routing engine, performance monitor and information exchange