[Ecn-sane] ect(1) queue selector question

Rodney W. Grimes 4bone at gndrsh.dnsmgr.net
Tue Mar 9 10:19:11 EST 2021


> I am of course much happier interacting here than over on tsvwg.
> 
> I would like very much
> 
> A large scale bit twiddling test from the larger providers of fq_codel
> and cake based hw and middleboxes.
> 
> extensive testing on lte and wifi transports. even emulated.
> the sce patches polished up and submitted for review to netdev as what
> we call there, an RFC
> the ability to test both SCE and L4S on openwrt (backport thereof)
> 
> I know there's been a lot of work on other qdiscs than cake, but
> haven't paid much attention. netdev review would be nice.
> 
> A simple internet-wide test of say bittorrent, or using an adserver
> method like what apnic uses.
> 
> Most of all I'd really like someone to take a stab at RFC3168 support
> for BBR. And by that, I don't mean a reduction by half, but to
> institute an RTT probe phase. A fixed rate reduction per RTT is simply
> not
> going to work IMHO, period, on lte and wifi. I'm sure dave miller
> would accept such a patch, and this would also lead towards better ecn
> support across the internet in general... at least from the
> perspective of the one provider I have an in with, dropbox.
> 
> SCE support for BBRv2 would be nice also.
> 
> And I know, a pony, all sparkly and stuff. I find it very difficult to
> summon the cojones to do a drop of work in this area, and I keep
> cheering you on - especially on the bug fixing and wonderful scripts
> and data you've provided so far, and I do wish I could find some way
> to help that didn't cause so much ptsd in me.
> 
> I wish I merely knew more about how dctp was configured in the data
> center. So far as *I* know its on dedicated switches mostly. I would
> vastly prefer a dscp codepoint for this, also.

If you want I can take a discussion on how DCTCP is configured in
data center switches in a side channel.  Its basically using a
drop function that has a starting queue depth, and slope that
as you go higher on the slope your propability of marking
increases.  The 40 gbit experiments we did with HPE modified
that part of the switch asic code to do SCE marks starting
at 1% mark probability for 1% Queue depth, up to 100% marking
at 100% Queue depth.  Sadly that was a slight mistake on
my part, and I thought we would get a chance to iterate
and retune these, I just pulled those values out as a WOG
and gave them to HPE to let them get on with setting it up.

For most practical purposes an SCE mark and a DCTCP CE
mark convey very similiar, if not the same information,
as does a CE mark in L4S.

Usually tunning of these queue depth and slope values involes
data collection and several iterations of experiments.

TOR vs SPINE switches are usually configured with different values.

I believe there is even some work by someone that does data
collection across the whole datacenter and tries to adjust
these automagically and dynamically.

> 
> On Mon, Mar 8, 2021 at 4:36 PM Pete Heist <pete at heistp.net> wrote:
> >
> > Sorry for reviving an old thread as I haven't been on this list in a
> > while:
> >
> > > > SCE proposes to use ect(1) as an indicator of some congestion and
> > does
> > > > not explictly
> > > > require a dscp codepoint in a FQ'd implementation.
> > >       Pretty much. I do think that a demonstration using an
> > > additional DSCP to create a similar HOV lane for SCE would have gone
> > > miles in convincing people in the WG that L4S might really not be as
> > > swell as its proponents argue, IMHO it won the day more with its
> > > attractive promise of low latency for all instead of what it
> > delivers.
> >
> > On that, I don't think any of us knows how things will end up or how
> > long it will take to get there...
> >
> > I do agree that the interim meeting leading up to the codepoint
> > decision could have gone better. Everything went great until it came to
> > how to deploy SCE in a small number of queues. We had dismissed the
> > idea of using DSCP, because we thought it would be panned for its poor
> > traversal over the Internet. That may still have been the case, but it
> > also may have worked if sold right. We thought that AF alone might be
> > enough to get past that part, but it wasn't.
> >
> > We already implemented a two-queue design that uses DSCP, but either
> > there wasn't much interest, or we didn't sell it enough. Plus, for
> > those who demand a two queue solution that requires no flow awareness
> > at all, DSCP alone may not get you there, because you still need some
> > reasonably fair way to schedule the two queues. So that might have been
> > the next line of criticism. Scheduling in proportion to the number of
> > flows each queue contains is one effective way to do that, but that
> > requires at least some concept of a flow. Perhaps there's another way
> > that doesn't introduce too much RTT unfairness, but I'm not sure.
> >
> > In our defense, there was already a lot of support built up for L4S,
> > and stepping in front of that was like stepping in front of a freight
> > train no matter what we did. I think we've made a decent argument in
> > the most recent version of the SCE draft that ECN is a "network
> > feature" which carries higher risks than drop-based signaling, and
> > warrants the requirement for unresponsive flow mitigation, for
> > starters. That of course requires some level of flow awareness, which
> > then makes various queueing designs possible. And, there may still be
> > deployment possibilities with DSCP, as Rodney mentioned.
> >
> > Anyway, there's progress being made on SCE, with some new ideas and
> > improvements to testing tools coming along.
> >
> > Pete
> >
> >
> > Ecn-sane at lists.bufferbloat.net
> dave at taht.net <Dave T?ht> CTO, TekLibre, LLC Tel: 1-831-435-0729
-- 
Rod Grimes                                                 rgrimes at freebsd.org


More information about the Ecn-sane mailing list