Hierarchical Congestion Control (HCC): Fairness and Fast Convergence for Data Centers

Congestion control protocols face several challenges for achieving max-min fairness and high throughput in a timely manner. In the absence of explicit signals, flows infer the congestion state solely based on their limited view of the network. Any attempt to expand the view with support from a centralized entity or network switches needs to deal with overheads as the number of flows scales. We present a system called Hierarchical Congestion Control (HCC) for data center networks. Flows that share bottlenecks are iteratively aggregated to build a hierarchy for propagating congestion control signals. HCC enables flow cooperation for sharing congestion information and collectively reacting to congestion. Aggregate control of multiple flows improves fairness and helps flows to converge faster in case of fluctuations in flow demands or network state. To limit the overheads, we compress congestion signals as they move up the hierarchy, pruning signals that do not involve significant changes. Our evaluation on real workload shows HCC converges to full utilization very fast with an empty queue and near-optimal fairness. Moreover, comparing to a centralized solution, HCC reduces the communication overhead to the core by 73% and the communication overhead from the core by 99.5%.