In steady state and single flow operation for 100% link utilization Reno requires a full BDP queue, CUBIC requires 40% BDP. This is independent of AQM or not. A tail-drop queue of the same size as CoDel's drop level performs exactly the same way in these circumstances. (I did the experiments.)
The benefit of AQM comes into play with many TCP flows in parallel. Then the queue size requirement goes down to something like BDP/sqrt(N) or even BDP/N (with N the number of flows). AQMs can earn this gain, tail-drop can not, it stays at BDP. For CoDel + CUBIC @100ms RTT that means, it's default 5ms target does not hurt, if more than 5 to 10 flows are present. (Taking into account that it drops slightly above the target.) Wolfram > For throughput to be unaffected with a single TCP-Reno stream the buffer > must grow to contain half the BDP, since the window will be halved on loss. > With TCP-CUBIC the buffer must contain 20% of the BDP. CoDEL will make > it's first drop when the sojourn time remains > 5ms, so with TCP-Reno, if > the RTT is > 10ms, then utilization will be hurt. With TCP-CUBIC if the RTT is > > 25ms then utilization will be hurt. > > Is this correct? > > Simon _______________________________________________ aqm mailing list aqm@ietf.org https://www.ietf.org/mailman/listinfo/aqm
