[Cake] Two questions re high speed congestion management and datagram protocols

[Cake] Two questions re high speed congestion management and datagram protocols

[David P. Reed]

[-- Attachment #1: Type: text/plain, Size: 1295 bytes --]


I was recently looking at congestion control algorithms - endpoint-based ones - that would deal with very low level, very low latency requirements in datacenters that use high speed switch fabrics. (note: congestion control in such datacenters is a very, very real issue, especially since some layer 2 switches are overbuffered by a huge factor, so they don't signal congestion while building very long queues at 40-400 Gb/sec)
 
I came across this work from MIT CSAIL [ https://ccp-project.github.io/ccp-guide/ ]( https://ccp-project.github.io/ccp-guide/ )
 
I also was looking back to DCCP as a useful way to get a UDP that handled congestion without engaging the higher layers, and preserving the other flexibility of UDP.
 
Anyone here have any experience with looking at the performance of these, especially w.r.t. Cake, which operates at the IP layer and thus takes direct advantage of IP congestion signalling? Does libreqos look like it might help?
 
[as the guy most associated with the creation of UDP, this remains an area of great interest personally, and is also presumably of relevance to QUIC... I don't know if either of these things are supported in systems other than Linux, but in server datacenters and my home lab, Linux networking is all that matters]
 
- David

[-- Attachment #2: Type: text/html, Size: 2443 bytes --]

Re: [Cake] [Bloat] Two questions re high speed congestion management and datagram protocols

[Stephen Hemminger]

On Sat, 24 Jun 2023 14:41:52 -0400 (EDT)
"David P. Reed via Bloat" <bloat@lists.bufferbloat.net> wrote:

> I also was looking back to DCCP as a useful way to get a UDP that handled congestion without engaging the higher layers, and preserving the other flexibility of UDP.

DCCP never got widely used, and Linux is on the path of deprecating it.