What you've said is true, and we will see the first order effect of that in our model. At that rate increase transition, user 1 will have ~20% of the link rate *available* to him. If he cannot capitalize on it because he has a small queue, then that capacity can be used by user 2 (who has a large queue). But, still, this effect on user 1 is only caused by user 1 himself in this case, user 2 has no impact. Comparatively speaking, it is true that user 2 would temporarily see a larger amount of free capacity on the link (> 20% available). This we don't try to model. I may be wrong, but I think this is a small nuance of a synthetic scenario.
Certainly, more complex models can be built to look into nuances of specific situations, and perhaps this is an area for other researchers to explore. But, at that point, I think you have to move away from simple synthetic usage models like we've been discussing in this thread (5 users simultaneously go immediately dark or simultaneously immediately saturate their link). TCP sessions (and other traffic loads) come and go across the user base asynchronously, so the available capacity fluctuates over time in a more random manner than we've used in this example. Similarly within each modem the traffic loads and TCP sessions will come and go and be in various states of congestion avoidance (hence differing amounts of queue), so the availability of a 500ms or 10ms of queued traffic isn't guaranteed in the FIFO vs FQ_CODEL cases. -Greg On 2/12/14, 1:15 AM, "Mikael Abrahamsson" <[email protected]> wrote: >On Tue, 11 Feb 2014, Greg White wrote: > >> What I was getting at is that the above is true for user #1 regardless >> of what applications and what AQM user #2 is using. So we don't need >>to >> simulate user #2 (it is sufficient to model him) if we're interested in >> user #1's performance. > >Are you sure? If AQM user #2 has 500 ms worth of packets in his upstream >buffer (FIFO) vs 10 ms worth of packets with FQ_CODEL, wouldn't that make >him adapt faster to increased upstream capacity? If nothing else, he has >plenty of packets buffered that can be emptied out, whereas with FQ_CODEL >you do not have this huge standing buffer with packets ready to send? > >-- >Mikael Abrahamsson email: [email protected] _______________________________________________ aqm mailing list [email protected] https://www.ietf.org/mailman/listinfo/aqm
