Re: [Dclc] Agenda topics

["Fred Baker (fred)" <[e-mail]>]

On Oct 27, 2014, at 4:54 PM, Bob Briscoe <[e-mail]> wrote:

> Fred,
> 
> At 23:34 27/10/2014, Fred Baker (fred) wrote:
> 
>> On Oct 27, 2014, at 3:33 PM, Bob Briscoe <[e-mail]> wrote:
>> 
>> > Fred, Lars,
>> >
>> > This is the same technique, but applied to a data centre by implementing the tunnel endpoints in the hypervisors:
>> > <https://tools.ietf.org/html/draft-briscoe-conex-data-centre-02>
>> >
>> > It's the one I had prepared a presentation for in the last dclc mtg, but I gave up the slot so we had longer to discuss the direction of the group.
>> > <http://www.bobbriscoe.net/presents/1407ietf/1407dc-congestion-policing.pdf>
>> > (I'd be happy to present it this time - perhaps in conjunction with Xinpeng if you establish that it will be relevant to data centres)
>> 
>> Maybe the two discussions. I agree that a vSwitch or hypervisor can be the tunnel endpoint, and RFC 2983 tells us how to manage the CU flags between the two IP headers. The question in my mind is how that interacts with latency.
> 
> I can't answer for Xinpeng, but in my case there's a congestion policer in each ingress hypervisor and the fill rates of all the tenants' token buckets (in units of ECN marks) bounds the growth of the FIFO AQM queue that all tenants share. This policer focuses drop on any tenant(s) that would otherwise cause heavier congestion, but it isolates all the other tenants from any queuing beyond the configured bound.
> 
> The pathetic part is I can't give any results, at least not in a data centre scenario.

I know the feeling. But we can discuss the concept.

> Bob
> 
> 
>> Where things stand right now, this is likely to be our entire agenda. There is some work from KAIST and RedHat/Morgan Stanley that I would have liked to see discussed, and they want to delay until March.
>> 
>> I'll post an agenda showing the two discussions. Happy to have discussion from the floor as well.
>> 
>> > And, to respond to Toby's IPR point, back in 2005 BT declared its IPR related to this, and updated the declaration in 2012, including royalty-free terms under the conditions specified here:
>> > <https://datatracker.ietf.org/ipr/1922/>
>> >
>> >
>> >
>> > Bob
>> >
>> > At 15:39 27/10/2014, Fred Baker (fred) wrote:
>> >> Content-Language: en-US
>> >> Content-Type: multipart/signed;
>> >>        boundary="Apple-Mail=_44D5A9BE-946A-4C98-96DA-75F726FDCBE8";
>> >>        protocol="application/pgp-signature"; micalg=pgp-sha1
>> >>
>> >>
>> >> On Oct 27, 2014, at 2:09 AM, Eggert, Lars <[e-mail]> wrote:
>> >>
>> >> > Hi Xinpeng,
>> >> >
>> >> > On 2014-10-27, at 09:21, Weixinpeng <[e-mail]> wrote:
>> >> >> It will be appreciated if there is a timeslot for me to show our work on tunnel-based congestion control.
>> >> >> The following is a link to the document.
>> >> >> http://datatracker.ietf.org/doc/draft-wei-tsvwg-tunnel-congestion-feedback/
>> >> >
>> >> > could you outline how this draft is related to latency control in datacenters? I read an earlier version and it seemed to be all about wide area networks?
>> >> >
>> >> > Thanks,
>> >> > Lars
>> >>
>> >> I'm scratching my head as well. Lingli is a co-author, so maybe she can comment here.
>> >>
>> >> The draft does mention NFV, which can be implemented across multiple data centers connected by a WAN, but if primarily implemented within a data center, and specifically mentions data centers. It also mentioned latency in passing, in the third paragraph of the section on 3GPP.
>> >>
>> >> Where I'm scratching my head is figuring out what it's actually trying to say. It explains NFV and that tunnels might be implemented in a vSwitch, which is true, and often the case in architectures such as OpenStack. It doesn't seem to make obvious points about that such as are made in RFC 2983: when the tunnel is marked as having experienced congestion, that information needs to be reflected in the interior header at the tunnel endpoint. It talks quite a bit about IPFIX. I'm not sure I walked away with "so do this".
>> >>
>> >>
>> >> _______________________________________________
>> >> Dclc mailing list
>> >> [e-mail]
>> >> https://www.irtf.org/mailman/listinfo/dclc
>> >
>> > ________________________________________________________________
>> > Bob Briscoe,                                                  BT
>> 
>> 
> 
> ________________________________________________________________
> Bob Briscoe,                                                  BT