CKCD: A Fair and Low Latency Queue Control Algorithm for Heterogeneous TCP Flows

The network congestion control algorithm BBR is proved to be effective in improving throughput and reducing latency. However, when BBR coexists with other loss-based congestion control algorithms (e.g., CUBIC, Reno), its flow unfairly consumes excessive bandwidth if the bottleneck node butter size is small. In this paper, we find that such inter-flow is caused by BBR's unresponsiveness to packet loss. To address this issue, we develop an in-network queue control algorithm called Choose-Keep and Controlled-Delay (CKCD), which actively penalizes flows that are either unresponsive to packet loss or have large queuing delay. CKCD requires limited packet information and hence is stateless and with low complexity. Extensive Mininet based experimental results show that, CKCD outperforms other recent benchmarks, including DropTail, RED, CHOKe and CoDel, with improved fairness and lower latency. In particular, compared to the widely adopted DropTail, CKCD improves inter-flow fairness by up to 62%, without deteriorating intra-flow fairness. It is also able to reduce the average latency by up to 92% while guaranteeing 96% link utilization.