查看文章

BBR is a model-based congestion control algorithm that has been widely adopted on the Internet. Different from loss-based algorithms, BBR features high throughput since it characterizes the underlying link and sends data accordingly. However, BBR suffers from high retransmission rates in deployment, leading to extra bandwidth costs. In this work, we carefully analyze and validate the reasons for high retransmissions in BBR flows. In a shallow-buffered link, the packet losses are deeply correlated to both the bottleneck buffer size and the in-flight data cap. Additionally, bandwidth drops also cause unwanted retransmissions. Based on the analysis, we design and implement oBBR, which aims at optimizing the retransmissions in BBR flows. In oBBR, we adaptively scale the in-flight data cap and update the bandwidth estimate timely so that few excessive data are injected into the network, avoiding packet losses. Our Internet experiments show that oBBR achieves 1.54× higher goodput than BBRv2 and 39.48% fewer retransmissions than BBR-S, which are both BBR variants with improved transmission performance. When deploying BBR in Internet streaming sessions, oBBR obtains greater QoE than BBRv2 and BBR-S without incurring more retransmissions. To summarize, oBBR is designed to help a transmission session reach high goodput and low retransmissions simultaneously, while other CCAs only achieve one of them.