Wolfram,
On 02/06/16 15:49, Lautenschlaeger, Wolfram (Nokia - DE) wrote:
Hi,
I don't think that the AQM story is completed right now. There is a number of
open issues that, I feel, are not well reflected by the currently discussed
AQM proposals.
Some examples:
- Bi-directional congestion (simultaneous up and download) produces a coarse
oscillation between the involved queues. Any heuristics on "standing" or
"steady" queue size should misinterpret that effect. Has it been investigated
for the actual AQMs?
- the "small RTT" issue (as Bob presented in Prague): At small RTT the AQM
might want to lower the CWND to less than 2 packets, which TCP does not
permit. As a result we get an uncontrolled queue growth despite AQM, ending up
with drastic drop rates >20%, tail drop, retransmission timeouts, etc. This is
a topic, not only, but also for AQM.
This should (at least at first) be dealt with in the various transport
protocol WGs that deal with TCP-like congestion controls (tcpm, on
behalf of mptcp, sctcp, etc), because it is caused by TCP ignoring drops
(or ECN) from the AQM when TCP's window ought to be less than 2 segments.
Nonetheless, it would be useful for the AQM WG to continue to exist in
order to review work on this.
Longer term, it is possible that a policer might be needed to detect and
drop packets from 'rogue' flows before they are allowed to increase the
delay of a shared queue. I am thinking of something like AFD that checks
packets randomly to determine potential rogue flows, rather than full
per-flow queuing (in which case there would be no shared queue to
protect). However, it would be pointless developing such a policer while
it would find that all standard TCP flows were 'rogue'. We need to
define a "better than rogue" behaviour for TCP first.
Up to now, AQM has avoided mission-creep into policing - perhaps that
should be reconsidered? Given, for addressing the queuing delay
problem, flow policing is more appropriate than flow queuing, and the
latter but not the former is included within the charter.
- degradation of per flow queuing in less-than-BDP buffers: AQMs are targeting
small queues, i.e. less than BDP. FQ is understood as the perfect complement
to achieve fairness. But at small buffers at a given moment only a fraction of
the involved flows have packets in their queue. Empty queues are not
participating in the round robin scheduling. The respective flows are not
earning credits as supposed for fairness reasons.
In general, this is a useful list of some new requirements, but
solutions (at least proposals) would be needed before AQM could really
get to grips with these.
(I'm trying to build a solution for the "small RTT" problem, but lots of
other demands on time.)
Bob
Just to mention a few of the open issues.
Wolfram
_______________________________________________
aqm mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/aqm
--
________________________________________________________________
Bob Briscoe http://bobbriscoe.net/
_______________________________________________
aqm mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/aqm
