查看文章

BBR is a new congestion-based congestion control algorithm proposed by Google. A BBR flow, if running alone, delivers the highest throughput with the lowest delay. To this end, each sender constantly measures the bottleneck bandwidth and round-trip propagation time of the network pipe, and uses the results to govern its behavior. However, our deployment in geo-distributed cloud servers reveals an unexpected RTT fairness problem: unlike conventional loss-based congestion control, BBR has an opposite bias towards long RTT flows. In this paper, we analyze the extent and cause of this fairness problem through cluster deployment in a clean network environment. Experimental results suggest that BBR’s bias against short RTT is an intrinsic problem of its design, irrespective of the bottleneck bandwidth, AQM deployment, and the number of coexisting flows. A BBR flow with 3× RTT of the other dominates with> 90% of bandwidth. Worse, strategic receivers can steal bandwidth by artificially inflating end-to-end delays. To address this problem, we show, through in-depth analysis, that a rapid-growing queue due to flow probing is the root cause of the bias. Based on this observation, we design a simple, yet effective algorithm, BBQ, with better RTT fairness while retaining BBR’s promising performance advantages. We have implemented BBQ in Linux kernel and evaluated it through extensive experiments. Our evaluation results show that BBQ improves RTT fairness by up to 6.1× in comparison to BBR. Moreover, BBQ reduces queueing delay by more than 64.5%, while maintaining nearoptimal link utilization.