ACC: Addressing Performance Limitations in Datacenters with Atomic Congestion Control

The proliferation of new datacenter applications drives the demand for ultra-low latency and high bandwidth underlying networks. However, we investigate that existing congestion control mechanisms suffer from bandwidth under-utilization due to overreaction or slowly ramp up when network perturbations occur. In this paper, we propose a novel window-based mechanism, called Atomic Congestion Control (ACC), to address the problems. During congestion, ACC directly decreases the rate of congested flows to their receiving rate to avoid overreaction and further leverages a corresponding sending window to drain out built-up queues. During the increase process, ACC combines the dynamic increase factor with the receiving rate to promptly reclaim free bandwidth. ACC ensures fairness and is easy to deploy in commodity datacenters. We theoretically analyze the convergence rate, fairness, and stability of ACC and further implement it with DPDK. Extensive testbed experiments and large-scale simulations show that ACC can cut the average queue length by up to 10× while reducing the average FCT by up to 82% compared with state-of-the-art approaches in large-scale simulations.