FastTune: Timely and Precise Congestion Control in Data Center Network
Renjie Zhou, Dezun Dong, Shan Huang, Yang Bai
Abstract
Modern data center networks (DCNs) exhibit high dynamics in both time and space dimensions, which poses challenges for congestion control protocols to achieve low latency, fast convergence, and high throughput. Existing methods have leveraged fine-grained link load information to achieve precise congestion control, but it still suffers from untimely control in highly dynamic DCNs. In this paper, we propose a timely and precise congestion control method called FastTune. FastTune employs fine-grained network status to achieve accurate feedback, uses switch feedback to control the first RTT, and leverages ACK-padding to shorten the feedback path and regulate congestion in time. FastTune develops a multiplicative increase/decrease (MI/MD) algorithm to achieve fast convergence based on timely and precise feedback. Large-scale evaluations show that, compared with state-of-the-art work, FastTune significantly reduces the feedback delay by up to 87%, reduces the average flow completion time by 40%, and the 99th percentile flow completion time by 51%. Besides, FastTune maintains near-zero queueing and reasonable throughput.