Chairs:
Jeff Tantsura (jefftant.ietf@gmail.com)
Yingzhen Qu (yingzhen.ietf@gmail.com)
WG Page: https://datatracker.ietf.org/group/rtgwg/about/
Materials: https://datatracker.ietf.org/meeting/120/session/rtgwg
##
Meeting Administrivia and WG Update
Chairs (10 mins)
https://datatracker.ietf.org/doc/draft-dmk-rtgwg-multisegment-sdwan/
Linda Dunbar (10 mins)
https://datatracker.ietf.org/doc/draft-acee-rtgwg-vrrp-rfc8347bis/
Acee Lindem (10 mins)
John asked a question in the chat (Sorry, John. The chairs missed the
queue.)
https://datatracker.ietf.org/doc/draft-ietf-rtgwg-arp-yang-model/
https://datatracker.ietf.org/doc/draft-zhang-rtgwg-ipv6-address-resolution-yang/
Fan Zhang (10 mins)
https://datatracker.ietf.org/doc/draft-yang-rtgwg-arn-framework/
Feng Yang (10 mins)
https://datatracker.ietf.org/doc/html/draft-liu-rtgwg-adaptive-routing-notification
Yao Liu
https://datatracker.ietf.org/doc/draft-zzhang-rtgwg-router-info/
Jeffrey Zhang
Jeff T: We've been getting drafts about fast notifications for a long
time, we will consider to have an interim focusing on this issue.
https://datatracker.ietf.org/doc/draft-cheng-rtgwg-adaptive-routing-framework/
Jiaming Ye
https://datatracker.ietf.org/doc/draft-he-huang-rtgwg-wan-lossless-framework/
Tao He (10 mins)
https://datatracker.ietf.org/doc/draft-xiong-rtgwg-use-cases-hp-wan/
Daniel Huang/Quan Xiong (10 mins)
Jeff T: There will be a side meeting on AIDC on Wednesday where some
similar problems will be talked about. There will be presentations about
adoption routing, congestion control as well.
The following presentations were not presented due to time limitation.
A OSF Framework for Artificial Intelligence (AI) Network
https://datatracker.ietf.org/doc/draft-hcl-rtgwg-osf-framework/
PengFei Huo (10 mins)
Use Cases-Standalone Service ID in Routing Network
https://datatracker.ietf.org/doc/draft-huang-rtgwg-us-standalone-sid/
Daniel Huang (10 mins)
If time permits:
Fast Congestion Notification Packet (CNP) in RoCEv2 Networks
https://datatracker.ietf.org/doc/draft-xiao-rtgwg-rocev2-fast-cnp/
Xiao Min
Loa Andersson
00:01:46
volume low on your mic
Yingzhen Qu
00:18:08
collective note taking: https://notes.ietf.org/notes-ietf-120-rtgwg
Ketan Talaulikar
00:20:10
I think the option 3 is better given the motivation. Suggest calling it
ietf-vrrp3 to reflect the protocol version which obsoletes the old RFC
Jim Guichard
00:20:40
someone is in the queue
John Scudder
00:21:38
I guess there was no time for Q/A? What I was in line to ask was if Acee
would Obsolete the v1 vrrp model, as well. I assume so but it wasn’t on
the slide.
John Scudder
00:21:48
Anyway seems reasonable.
Yingzhen Qu
00:22:18
John, can you please send your question to the list?
Jim Guichard
00:33:03
@chairs please keep an eye on the queue
Yingzhen Qu
00:33:37
@Jim, sure.
David Black
00:34:03
2B = To Business?
Daniel Bernier
00:44:37
If would be interesting to validate the challenge versus BBF work with
traffic steering
Daniel Bernier
00:44:43
ie WT-474
Daniel Bernier
00:44:51
that is basically the purpose of it
Jeff Tantsura
00:45:31
Dan - could you please send an email to the list with your comment?
Daniel Bernier
00:45:44
yes I will ;-)
Joel Halpern
00:51:03
I saw no reference in this draft to RFC 2386. Nor to the hard problems
that the RFC discusses. Without that, I do not see the point of
disucssing a notification protocol.
Joel Halpern
00:51:33
(Sent to chat since the chairs do not seem to want discussion at the
microphone.)
Tony Li
00:58:49
Loop avoidance is the responsibility of your PCE
Nitsan Dolev Elfassy
01:00:32
PCE may not be able to follow the changes ...
Tony Li
01:02:26
If the changes are at too high a frequency for a PCE, then you lack a
stable control loop.
Boris Khasanov
01:02:27
If PCE gets LSDB (TED) it will but with some delay
Joel Halpern
01:03:53
The whole approach lacks a stable control loop. THe authors may or may
not have one in mind, but the congestion response draft doesn't include
it.
Liu Yao
01:08:42
@Joel so the suggestion is to add the control loop in the draft on how
to response to the congestion notification, right?
Joel Halpern
01:12:09
@liu Yao No. My suggestion is to start by explaining why the problems
identified in the mid-90s with any approach like this don't apply. Only
after those problems are recognized and discussed is there any point in
discussing mechanisms.
Nitsan Dolev Elfassy
01:12:47
The whole approach does not analyze the load of the response traffic
itself ...
Joel Halpern
01:12:58
It may be that you can find a way around those problems. But that is
where you need to start.
Liu Yao
01:14:55
Got it, will work on this ,thanks for pointing it out
David Black
01:23:08
Is this mechanism capable of re-ordering packets within a flow?
Liu Yao
01:23:15
@Nitsan generally, ARN suppression mechanism should be used accordingly
to avoid notification storm. will add the analyze on this aspect.
Tony Li
01:23:35
Not just capable...
Jeff Tantsura
01:24:53
it will cause packet reordering
David Black
01:25:08
Not good for a number of higher-layer protocols.
Tony Li
01:25:22
The only upside is that it will be rate limited by BGP convergence time.
David Black
01:25:34
Packet-based adjustment mode looks like a particular problem.
Joel Halpern
01:26:14
Moving an elephant flow does seem highly likely to simply move the
congestion, not avoid it.
Jeff Tantsura
01:26:46
in DC (10ns RTT) - BGP as the main control loop signaling is not very
suitable for that
David Black
01:26:48
And if the elephant transport stumbles due to reordering around the
move, the move appears to help ... until it doesn't.
Nitsan Dolev Elfassy
01:28:10
Should not one expect an analysis of the various congestion causing
factors before approaching with these congestion control tools ? Have we
done such analysis ?
Cheng Li
01:29:42
I would like to know how terrible caused by congestion today, before
discussing the solution
Dan Voyer
01:30:03
good point Cheng
John Scudder
01:30:40
If only there were some prior art on load-sensitive routing.
Joel Halpern
01:31:41
While I presume some of the reality has changed, RFFC 2386 does seem a
good starting point :-)
Tony Li
01:31:50
? Congestion comes from having insufficient bandwidth available to
transport the available load. This may be due to insufficient capital,
insufficient deployment, or inadquate ECMP paths. For DC purposes, a
good first order approximation is an infinite demand.
Eduard V
01:32:02
Cheng, ping me - I will explain. It is a consequence of cutting upper
layer switches from the hyper-scale DC/Cloud (replacing by mesh).
Cheng Li
01:32:31
thank you Eduard :)
Dan Voyer
01:33:01
@tony Li, it could be traffic spike - ex.: when the game industry
release a new path/upgrades it does create spike
Tony Li
01:34:45
@Jeff Tantsura I agree that any current control plane is out of the
question. One could, in principle have faster, lower-level control
loops.
Joel Halpern
01:35:34
One could congest two paths at once by using detnet preof :-)
Jeff Tantsura
01:35:44
@tli this is the topic of most today' presentations, some kind of "fast
notification"
Tony Li
01:35:47
@Dan Voyer You're absolutely right, I left out insufficient foresight
and planning.
Weiqiang Cheng
01:38:04
Jeff, I agree with you that BGP maybe not suitable, and we need some
fast solution. It is possible because the AIDC network topo is simpler
Jeff Tantsura
01:41:45
Important point to consider - perhaps transport network is not the right
place to solve the problem; end-points in most cases run some kind of
congestion control, either ECN marking based (e.g DCQCN) or timer based
(e.g. timely) - these are much faster in detection and can mitigate at
the source - by reducing TX rate or changing entropy
Weiqiang Cheng
01:41:47
@cheng li, the adaptive routing is designed for AIDC, in which the
traffic model is really different from IP WAN network. The size per flow
may be 400G for such as all-reduce traffic from only one GPU.
Cheng Li
01:44:38
it is huge, so it might be worthy to investigate :)
Tony Li
01:52:12
Shouldn't this be in QUIC and answered by same?
David Black
01:53:05
Or tsvwg fot the CC aspects. Dan's mike comment on relationship to
detnet is very much on point.
Weiqiang Cheng
01:53:07
@Dan, I don't think IP WAN network need adaptive routing. It is for AI
training data center. When you run training task, at the same time, all
the GPU involved in one all-reduce task maybe running at line rate, such
as 400Gbps. It will cause significant congestion, and then lower the
training efficiency. The traffic model is much worse than that in IP WAN
network even in worst case.
Jeff Tantsura
01:54:02
in CCWG we are progressing HPCC documents, these are INT based (switch
assisted) congestion control for AIDC