DWTCP: Ultra Low Latency Congestion Control Protocol for Data Centers

Abstract

Congestion control algorithms rely on a variety of congestion signals (packet loss, Explicit Congestion Notification, delay, etc.) to achieve fast convergence, high utilization, and fairness among flows. A key limitation of these congestion signals is that they are either late in feedback or they incur significant overheads. An ideal congestion control must discover any available bandwidth in the network, detect congestion as soon as link utilization approaches full capacity, and react timely to avoid queuing and packet drops, without significant overheads. To this end, this work proposes Scout service that leverages priority queues to infer bandwidth availability and link busyness at the host. The key observation here is that as the high priority queue (HPQ) gets busier, the low priority queue (LPQ) is served less. Therefore, the state of the link can be observed from the LPQ and any congestion can be detected several RTTs earlier than observing the HPQ. We propose a new transport protocol, Double-Window Transmission Control Protocol (DWTCP) that builds upon the Scout service to dynamically adjust its congestion window. Our testbed and simulation-based evaluation demonstrates that Scout enables a data center transport to achieve high throughput, near-zero queues, lower latency, and high fairness.