A unified approach to optimizing performance in networks serving heterogeneous flows

We study the optimal control of communication networks in the presence of heterogeneous traffic requirements. Specifically, we distinguish the flows into two crucial classes: inelastic for modeling high-priority, delay-sensitive, and fixed-throughput applications; and elastic for modeling low-priority, delay-tolerant, and throughput-greedy applications. We note that the coexistence of such diverse flows creates complex interactions at multiple levels (e.g., flow and packet levels), which prevent the use of earlier design approaches that dominantly assume homogeneous traffic. In this work, we develop the mathematical framework and novel design methodologies needed to support such heterogeneous requirements and propose provably optimal network algorithms that account for the multilevel interactions between the flows. To that end, we first formulate a network optimization problem that incorporates the above throughput and service prioritization requirements of the two traffic types. We, then develop a distributed joint load-balancing and congestion control algorithm that achieves the dual goal of maximizing the aggregate utility gained by the elastic flows while satisfying the fixed throughput and prioritization requirements of the inelastic flows. Next, we extend our joint algorithm in two ways to further improve its performance: in delay through a virtual queue implementation with minimal throughput degradation and in utilization by allowing for dynamic multipath routing for elastic flows. A unique characteristic of our proposed dynamic routing solution is the novel two-stage queueing architecture it introduces to satisfy the service prioritization requirement.