Confucius Queue Management: Be Fair But Not Too Fast

When many users and unique applications share a congested edge link (e.g., a home network), everyone wants their own application to continue to perform well despite contention over network resources. Traditionally, network engineers have focused on fairness as the key objective to ensure that competing applications are equitably and led by the switch, and hence have deployed fair queueing mechanisms. However, for many network workloads today, strict fairness is directly at odds with equitable application performance. Real-time streaming applications, such as videoconferencing, suffer the most when network performance is volatile (with delay spikes or sudden and dramatic drops in throughput). Unfortunately, "fair" queueing mechanisms lead to extremely volatile network behavior in the presence of bursty and multi-flow applications such as Web traffic. When a sudden burst of new data arrives, fair queueing algorithms rapidly shift resources away from incumbent flows, leading to severe stalls in real-time applications. In this paper, we present Confucius, the first practical queue management scheme to effectively balance fairness against volatility, providing performance outcomes that benefit all applications sharing the contended link. Confucius outperforms realistic queueing schemes by protecting the real-time streaming flows from stalls in competing with more than 95% of websites. Importantly, Confucius does not assume the collaboration of end-hosts, nor does it require manual parameter tuning to achieve good performance.