Updating the Theory of Buffer Sizing

Abstract

Routers have packet buffers to reduce packet drops during times of congestion. It is important to correctly size the buffer: make it too small, and packets are dropped unnecessarily and the link may be underutilized; make it too big, and packets may wait for a long time, and the router itself may be more expensive to build. Despite its importance, there are few guidelines for picking the buffer size. The two most well-known rules only apply to long-lived TCP Reno flows; either for a network carrying a single TCP Reno flow (the buffer size should equal the bandwidth-delay product, or $BDP$) or for a network carrying $n$ TCP Reno flows (the buffer size should equal $BDP/\sqrt{n}$). Since these rules were introduced, TCP Reno has been replaced by newer algorithms as the default congestion control algorithm in all major operating systems, yet little has been written about how the rules need to change. This paper revisits both rules. For the single flow case, we generalize the $BDP$ rule to account for changes to TCP, such as Proportional Rate Reduction (PRR), and the introduction of new algorithms including Cubic and BBR. We find that buffers can be made 60-75% smaller for newer algorithms. For the multiple flow case, we show that the square root of $n$ rule holds under a broader set of assumptions than previously known, including for these new congestion control algorithms. We also demonstrate situations where the square root of $n$ rule does not hold, including for unfair flows and certain settings with ECN. We validate our results by precisely measuring the time series of buffer occupancy in a real network, and comparing it to the per-packet window size.