TCP Cubic

Abstract

TCP Cubic changes the congestion control algorithm. During congestion avoidance, increases congestion window faster than TCP Reno on fast long-distance networks → Slows rate of increase as window approaches the largest successfully achieved window

Ex: On packet loss during congestion avoidance, TCP Cubic reduces congestion window: $W_i=W_{i-1}\times 0.7$ TCP Reno uses $W_i=W_i-1\times 0.5$ TCP Cubic is more aggressive

After packet loss during congestion avoidance, TCP Cubic increases the congestion window as: $W_{cubic}=C\ast (t-K{)}^3+W_{max}$

$W_{max}$ is the maximum window size reached before the loss • $t$ is the time since the packet loss • $K$ is the time it will take to increase the window back to $W_{max}$, assuming no further packet losses • $C=0.4$ is a constant that controls fairness to TCP Reno

Many additional details included to ensure fairness with TCP Reno on slower, shorter-RTT, networks

TCP Cubic is default in most modern operating systems

• Much more complex than TCP Reno – core response is relatively straight-forward
• Much complexity ensures fairness with TCP Reno in its typical operating regime; improves performance for networks with longer RTT and higher bandwidth Both algorithms use packet loss as congestion signal and eventually fill router buffers Trade latency for throughput