A Principled Look at the Utility of Feedback in Congestion Control

Networked applications are ubiquitous and their performance requirements are becoming increasingly stringent. Network congestion can seriously impact performance contributing to increased latency, packet loss and poor throughput. To address these problems, the networking community has come up with a large number of congestion control algorithms. Congestion control schemes developed over the past few decades can be classified into two broad classes: one based on an end-system's perception of network congestion and the other based on the network providing feedback to flows that pass through it. In this paper, we make the observation that the pure end-system based congestion control schemes are faced with the significant challenge of receiving ambiguous signals that make it difficult to infer where the congestion is occurring and if this flow is even the cause of that congestion. This ambiguity makes it difficult for pure end-system based control schemes to achieve fairness across different flows. Modern routers and switches in the meantime, have grown in computing capability and can generate fine grained feedback at line speeds for flows traversing them. We show that even relatively simple feedback generated in-network at the point of congestion eliminates the ambiguities faced by pure end-system based congestion control mechanisms, thus ensuring the network functions at the right fair and efficient operating point. We provide the theoretical underpinnings establishing the need for in-network feedback to enable the network to operate at a unique fixed point at the intersection of the desired fair and efficient operation regimes, and demonstrate through emulation experiments that our use of the well-established and studied PI-control for Active Queue Management and Explicit Congestion Notification meets the goals of low latency, high throughput and fine granularity control of the queue while achieving fairness.