Wes,
My 2 pennies worth inline
At 02:56 26/11/2012, Wesley Eddy wrote:
As TSV ADs, Martin and I have been wondering lately about
AQM work.
At this point, thanks largely to Jim Gettys, we all know
about large buffers, and we know what bulk-transfer with
loss-based congestion control can do to interactive traffic
when large queues exist.
We know one thing that can be done to help is the use of
AQMs, but the IETF is currently not doing a whole lot to
help in this regard.
There is the CoDel I-D that was discussed in the TSVAREA
meeting in Vancouver:
http://tools.ietf.org/html/draft-nichols-tsvwg-codel-00
But this is an individual submission that Martin and I
would be happy to sponsor, and not a working group
activity.
At the last IETF meeting, we had a request to talk about
PIE, which we didn't have time for in TSVAREA, but ICCRG
did:
http://www.ietf.org/proceedings/85/slides/slides-85-iccrg-2.ppt
This was really interesting, and we wonder if the
community is interested in continuing to hear about
AQMs in this space, and what the IETF should be doing.
If people have thoughts about this, we'd like to hear
about them on this list, and/or during the TSVAREA
meeting in Orlando at the next IETF meeting. We want
to know what the level of interest in AQMs is, and
might make this a focus of the TSVAREA meeting, if a
lot of people seem to think it's a good idea.
Particularly, I have wondered:
- Should we have a working group looking at AQMs?
Yes. During the lead-up to the RMCAT workshop & BoF, a number of us
said that there are three problems for r-t apps, which I would rank
in this order of importance (highest first):
- AQM behaviour
- competing TCP congestion avoidance behaviour
- r-t app congestion avoidance behaviour
So it makes sense to tackle the most important one as well as the
least important.
- If so, does it make sense to shoot for Standards Track
specifications?
Well, I would have thought informational specs of various algos
proven to work well
and a BCP saying what any algo should / should-not do.
- Would we be able to come up with actual requirements
on an AQM so that it's implementable for cheap in
hardware and software, and behaves well under load?
Yes
- Would it be valuable to have a test suite for AQMs
similar to what ICCRG was doing for high-rate TCP?
Yes and no.
No, in that the tests for an AQM should be tailored to find
deficiencies in a specific AQM that may be suspected by understanding
the design of each one. Certainly, after that, other AQMs should be
subjected to similar tests. However, this implies that the test
surface should continually expand. This extensibility could be
described in an RFC, but it implies a living document is also required.
Yes, in that we can't expect everyone to think from first principles,
so a process people can use that saves them thinking can be useful.
- Would it be valuable to have a BCP (or multiple) on
configuring "legacy AQM" like RED, or the use of AQM
metrics like "sojourn time" rather than variations of
the queue length?
Yes. The advice in RFC3819 needs a serious update, for instance.
Recently an industry consultant asked me where there is good advice
on how to config existing switches and routers, and I realised the
level of advice on the Web, let alone in the RFCs, is well out of date.
- If any of this is of value, how much belongs in the
ICCRG versus TSV working groups?
Well, RED and particularly WRED is very widely deployed. So that
implies at least guidelines about legacy should be in TSV not ICCRG.
We are really gunning for self-configuring AQMs. They are not yet
widely deployed so one could say that is a research subject. However,
static config AQMs have been researched to death already (and
parameter sensitivity), so I think we can say the area in general is
understood enough to go straight into the IETF, not ICCRG.
AQM for wireless interfaces has been less researched - probably
better to have ICCRG activity in parallel, but allow anything
sufficiently mature in the IETF too.
Joe Touch says INT or OPS. I don't agree - there isn't the expertise
there. I admit the core competence in TSV is e2e protocols, but TSV
is where nearly all the expertise in queuing sits too. We should
present to INT & OPS, but not do the work there.
Thanks for raising this.
Bob
Thanks for your thoughts :).
--
Wes Eddy
MTI Systems
________________________________________________________________
Bob Briscoe, BT Innovate & Design
