AdaptaCC: A Highly Scalable Proactive Congestion Control for Hyper-speed Networks.

Emerging applications such as 4K/8K UHD video, holographic communication and large-scale AI training are extremely dependent on hyper-speed transmission of massive data. Legacy TCP suffers from low throughput due to its conservative congestion control mechanism, while the high overhead of the terminal and the shallow buffer of the switch in hyper-speed networks also become the new performance bottlenecks. Emerging proactive congestion control protocols support hyper-speed transmission by quickly allocating link bandwidth between active flows, but existing protocols require extensive modifications to the system kernel and significantly increase the overhead of the switches due to packet-by-packet processing, making them impractical for deployment. In addition, almost all existing protocols cannot achieve bandwidth-independent high scalability and have the potential to be applied in future networks with more than Tbps bandwidth. In this paper, we propose a novel highly scalable proactive congestion control for hyper-speed networks, called AdaptaCC. AdaptaCC implements transmission control through a few signal packets, enabling fast, efficient and fair link bandwidth allocation, as well as low processing overhead and high scalability. In addition, AdaptaCC rapidly perceives congestion on the bottleneck link and precisely adjusts the transmission rate though the RTT difference. Extensive simulation results show that compared with the baseline protocols, AdaptaCC can quickly converge to utilize link bandwidth while avoiding congestion, and it shortens flow convergence time by up to 87%. AdaptaCC can maintain near-zero in-network queues, and it reduces interference recovery time by nearly 53%. Moreover, AdaptaCC is the only protocol that can support 400Gbps transport in NS2.