[Ecn-sane] FQ in the core

Sebastian Moeller moeller0 at gmx.de
Mon Mar 25 05:52:10 EDT 2019


Hi Luca.

> On Mar 25, 2019, at 10:17, Luca Muscariello <luca.muscariello at gmail.com> wrote:
> 
> We've had this discussion multiple times.
> 
> - You do not need FQ everywhere and depending on 
> the case you can do approximations of that.
> 
> - What I believe you always need is flow-awareness.
> 
> - There are already implementations of dual-queue systems in DC switches such as the Cisco nexus 9k.
> 
> - The dualQ system in docsis is the wrong way to implement a flow-aware system with two queues. 
> For many reasons including, but not limited to the fact  that dualQ to work makes assumptions about the behaviour of the end-points.
> A flow aware queuing  system should not be sensitive to some sort of compliancy of the end-points.

+1


> You cannot trust the end-points and the protection system should not discriminate good vs bad based on a badge carried by the packets.

Exactly my sentiment, "speak softly, and carry a big stick" or "Doveryai, no proveryai" in any way do not give special treatment just because someone asks nicely, in other words "Pedo mellon a minno" considered harmful.


> 
> - I also believe  dualQ fails to achieve its goal under current and future traffic patters.

	I would very much like to test that, maybe the VMs intended as one of the hackathon's goals will make playing with that easier.

> 
> - The approach described in this paper also works for 2 queues only and makes no assumption about the end-points.
> It does not require marking at all.
> 
> James Roberts et al.
> Minimizing the overhead in implementing flow-aware networking. 
> In Proceedings of the 2005 ACM symposium on Architecture for networking and communications systems (ANCS '05). 
> DOI: https://doi.org/10.1145/1095890.1095912
> https://team.inria.fr/rap/files/2013/12/KMOR05a.pdf
> 
> -  There is not one TCP, there is no one single transport in the wild. There are many and there will be many more.
> The docsis specs makes the assumption that to be a good guy you must be part of one Church. The one specified.
> This is a very religious assumption about end-points' behaviour.
> All the others are bad guys. I'm the only one who sees this as deeply wrong?

	I tried to make this point in the past, LLLLS seems to accidentally position itself as the be-all end-all of internet transport and if we know one thing about the future it is its crappy predictability in the long run...

> 
> - Almost 10 years ago I built a FQ prototype on an NPU based on Cavium with Alcatel-Lucent (the team was based in Paris). 
> That was supposed to be an ALU7750 target. The problem was that not a single ISP was even aware of the topic. Nobody was asking for it.
> It's a chicken and egg problem Mikael. If you do not ask loudly nobody builds it.

	How could something like this be priced? On a port by port basis, like vectoring for VDSL2, then ISPs could still offer something like that as an extra they could sell for latency sensitive customers?


> 
> - I also built with Ikanos in 2010, a FQ prototype in their MIPS based SoC. 32 queues for the France Telecom livebox
> and it worked. Ikanos was later acquired by Qualcomm and that SoC is not used anymore in favour of Broadcom.

	Ah, this is why ikanos disappeared...


> 
> Hope this helps to make progress in the discussion

	All, quite interesting, gracie!

Best Regards
	Sebastian

> 
> Luca
> 
> 
> 
> 
> 
> On Mon, Mar 25, 2019 at 8:55 AM Dave Taht <dave.taht at gmail.com> wrote:
> I don't really have time to debate this today.
> 
> Since you forked this conversation back to FQ I need to state a few things.
> 
> 1) SCE is (we think) compatible with existing single queue AQMs. CE
> should not be exerted in this case, just drop. Note that this is also
> what L4S wants to do with the "normal" queue (I refuse to call it
> classic).
> 
> 2) SCE is optional. A transport that has a more agressive behavior,
> like dctcp, should fall back to being tcp-friendly if it
> sees no SCE marks and only CE or drop.
> 
> 3) At 100Gbit speeds some form of multi-queue oft seems needed. (and
> this is in part why folk want to relax ordering requirements). So some
> form of multiple queuing is generally the case. At the higher speeds,
> DC's usually overprovision anyway.
> 
> 4) The biggest cpu overhead for any of this stuff is per-tenant (in
> the dc) or per customer shaping. This benefits a lot from a hardware
> assist. (see senic). I've done quite a bit of DC work in the past 2
> years (rather than home routers), and have had a hard look at the
> underlying substrates for a few multi-tenant implementations....
> 
> 4) "dualq" hasn't tried to address the fact that most 10Gbit and
> higher cards have 8 or more hardware queues in the first place.
> 
> 5) Companies like preseem are shipping transparent bridges that do
> fq_codel/cake on customer traffic.
> 
> I've long been in periodic negotions with makers of "big iron" like,
> for example, the new 128 core huwei box and others I cannot talk about
> at the moment, to get so far as an existence proof.
> 
> So I'd like to kill the meme that SCE requires FQ, at least, for now,
> until after we do more tests.
> 
> As for FQ everywhere, well, I'd like that, but it's not needed in
> devices that already have sufficient multiplexing.
> 
> 
> 
> 
> 
> On Mon, Mar 25, 2019 at 8:16 AM Mikael Abrahamsson <swmike at swm.pp.se> wrote:
> >
> > On Sun, 24 Mar 2019, Sebastian Moeller wrote:
> >
> > > From my layman's perspective this is the the killer argument against the
> > > dualQ approach and for fair-queueing, IMHO only fq will be able to
> >
> > Do people on this email list think we're trying to trick you when we're
> > saying that FQ won't be available anytime soon on a lot of platforms that
> > need this kind of AQM?
> >
> > Since there is always demand for implementations, can we get an ASIC/NPU
> > implementation of FQ_CODEL done by someone who claims it's no problem?
> >
> > Personally I believe we need both. FQ is obviously superior to anything
> > else most of the time, but FQ is not making itself into the kind of
> > devices it needs to get into for the bufferbloat situation to improve, so
> > now what?
> >
> > Claiming to have a superior solution that is too expensive to go into
> > relevant devices, is that proposal still relevant as an alternative to a
> > different solution that actually is making itself into silicon?
> >
> > Again, FQ superior, but what what good is it if it's not being used?
> >
> > We need to have this discussion and come up with a joint understanding of
> > the world, otherwise we're never going to get anywhere.
> >
> > --
> > Mikael Abrahamsson    email: swmike at swm.pp.se
> > _______________________________________________
> > Ecn-sane mailing list
> > Ecn-sane at lists.bufferbloat.net
> > https://lists.bufferbloat.net/listinfo/ecn-sane
> 
> 
> 
> -- 
> 
> Dave Täht
> CTO, TekLibre, LLC
> http://www.teklibre.com
> Tel: 1-831-205-9740
> _______________________________________________
> Ecn-sane mailing list
> Ecn-sane at lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/ecn-sane
> _______________________________________________
> Ecn-sane mailing list
> Ecn-sane at lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/ecn-sane



More information about the Ecn-sane mailing list