On 29/08/2014 14:52, David Collier-Brown wrote:
On 08/29/2014 09:16 AM, Scheffenegger, Richard wrote:
Hi Gorry,
Given QUIC includes FEC to hide losses, I guess it is a good example to
consider whether ECN still offers sufficient benefits over and above
just removing losses.
GF: And then, isn't the implication of AQM to significantly increase the
number of "losses" unless we use ECN?
Indeed, I have the impression we are confusing many on these points -
ECN could change the reaction to congestion signal, and FEC (even
opportunistic CC-friendly FEC) can also change the way things react to
congestion signals.
I don't think that an AQM's implication is automatically to increase the number
of losses; that may happen to specific flows (in particular, unresponsive
ones), but for responsive (non-ECN) ones, the expectation would be to
de-correlate the losses, and for TCP, to only have around 1 loss per window
when necessary - instead of a burst loss of one window and the expensive
recovery...
Perhaps it's that perception that also poses an obstacle to AQM deployment,
because of the believe that a dynamic but lower mark/drop threshold will cause
more losses?
Goodness gracious, from the point of view of a queuing network, AQM
reduces losses overall, in the process of minimizing delay and keeping
bandwidth use just below the theoretical maximum.
Oversimplifying, we try to keep the buffers empty, so that if we get a
burst we can handle it without losses and without affecting other
communications. We signal via a loss or other indicator if the
non-bursty flow is enough to cause congestion, which keeps the buffers
near-empty and the system uncongested.
Yep that's what I think is the goal.
Failing to do so fills the buffers without signalling there is
congestion, and induces delay on everyone who's dependant on the buffer.
Not to mention allowing the congestion to go unreported!
It's not a tradeoff discussion: it's arguably one about correctness.
--dave
[Somebody like Neil Gunther could explain the math of this better, but
the behaviour is well-known in the trade, and cordially hated.
Congestion control is superior to admission control, which is what I
often use to prevent the server equivalent of congestive collapse (:-))]
I agree that AQM often reduces losses (what I said above didn't say
this) but it sometimes it can increase loss (e.g. a large RTT or less
responsive CC), but I don't recall saying something on this in the
draft. If you see something in the draft that isn't correct then let us
know.
Gorry
_______________________________________________
aqm mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/aqm
