Hi Dave and all, On 07.07.2016 at 21:11 Dave Täht wrote: > having RTOs? I do not think that a fixed delay is desirable. "You aim > for a target", not that you need to hit it. More important is high > utilization, and trying to keep the connections afloat, with as minimum > latency as possible.
That is correct. There are cases where a fixed target is sub-optimal, because one can achieve still full utilization with a lower target. We just submitted a paper discussing several interactions of TCP congestion control and AQMs. On the one hand results from experiments with Codel, PIE and GSP showed that in cases with several flows, AQMs can achieve good utilization depending on how well they achieve loss-desynchronization. We reduced the threshold of GSP by more than 50% and still achieved high utilization. On the other hand, for single flows or only a few flows, AQMs cannot achieve full utilization, because TCPs window reduction is too large in most cases. So it's more like having an upper bound for queueing delay than a target and the stable point of operation may be well below that bound. Regards, Roland _______________________________________________ aqm mailing list aqm@ietf.org https://www.ietf.org/mailman/listinfo/aqm
