Re: [tsvwg] a new method of congestion control

["Shihang(Vincent)" <[e-mail]>]

From https://nvidianews.nvidia.com/news/nvidia-announces-spectrum-high-performance-data-center-networking-infrastructure-platform :
With 51.2Tbps aggregate ASIC bandwidth supporting 128 ports of 400GbE, combined with adaptive routing and enhanced congestion control mechanisms, Spectrum-4 optimizes RDMA over Converged Ethernet fabrics and dramatically accelerates data centers.

AI workloads run mostly on RoCE, not TCP.

Best,
Hang

-----Original Message-----
From: tsvwg <[e-mail]> On Behalf Of Mitchell Erblich
Sent: Monday, June 5, 2023 12:14 PM
To: [e-mail]
Cc: tsvwg IETF list <[e-mail]>
Subject: Re: [tsvwg] a new method of congestion control

Joe,

	One would need to duplicate Appropriate Byte Counting (ABC) by Allman / Pacing in Ethernet, but why would you want to when you already had it within a experimental CA in TCP?

	If you flood packets and a intermediate system (router) did a tail-drop or a random-early drop, you still drop packets.

	If you have out-of-orderness that jumps because… then you drop packets

	If you fragment and drop a fraction of the packet you drop…
	
	If you re-route because of the change of the routing table and you re-order, you drop,,,,

	Yes,we have selective asks…. But not in Ethernet or IPvX.

	Ethernet supports two sizes of IP (IPv4 and IPv6) … do you do something special for a fraction of the packets in your network?

	You need to have 1 hop networks … Again ,,, weird…

	Can’t guarantee no drop of packets… Yes, you can mitigate if and if  and if… but is’t it just worth to use Selective Acks and other existing RFCs 

	Why create a wheel when it already exists???


Group,,, Sorry, about the verbose response….

Mitchell Erblich
[e-mail]



> On Jun 4, 2023, at 8:15 PM, [e-mail] wrote:
> 
> Ethernet has a LOT of extensions:
> 
> IEEE 802.1
> en.wikipedia.org
> <wikipedia.png>
> 
> Could this either be some subset of newer extensions or revival/vendor-specific variants of past attempts?
> 
> Joe
> 
> —
> Dr. Joe Touch, temporal epistemologist www.strayalpha.com
> 
>> On Jun 4, 2023, at 7:50 PM, Mitchell Erblich <[e-mail]> wrote:
>> 
>> Weird, In my Opinion (IMO),
>> 
>> 	How do you separate out AI versus non-AI flows, so you can tailer one type network flow to AI aware applications?
>> 
>> 	Is AI consuming even 20% of network bandwidth at this time?
>> 
>> 	Isn’t Apple’s ?Siri? an AI application approaching 10 years of age, which I find far from perfect.
>> 
>> 	It is like having a different Ethernet protocol for UDP versus TCP.
>> 
>> 	The article must be missing critical information or the writer doesn’t understand TCP and/or CA.
>> 
>> Mitchell Erblich
>> [e-mail]
>> 
>> 
>> 
>>> On Jun 4, 2023, at 5:53 PM, Dave Taht <[e-mail]> wrote:
>>> 
>>> announced by nvidia here:
>>> https://www.zdnet.com/article/nvidia-unveils-new-kind-of-ethernet-fo
>>> r-ai-grace-hopper-superchip-in-full-production/
>>> 
>>> I have no idea what an AI workload looks like, familiarity with a 
>>> ton of DC l2 protocols, and there are hints in this post about telemetry.
>>> 
>>> Anyone have clue here?
>>> 
>>> --
>>> Podcast: 
>>> https://www.linkedin.com/feed/update/urn:li:activity:705879391022711
>>> 1937/
>>> Dave Täht CSO, LibreQos
>>> 
>> 
>