Re: [L4s-discuss] rpi 5 and hardware ptp

[rjmcmahon <[e-mail]>]

On 2024-04-08 13:57, Rodney W. Grimes wrote:
>> >
>> > Your kernel should have full support for ECN and PACING.
>> > sysctl net.ipv4.tcp_ecn
>> > to check for ecn state, 2 is the default, you probably want to
>> > flip this to 1 on the end running iperf -s to get a ecn connection.
>> >
>> > PACING is a setsockopt, try man tc or man fq and search for pacing.
>> >
>> 
>> Thanks for this. Which CCA will impelement ECN based pacing?
> 
> I know of none that implement ECN and Pacing at the same time,
> other than L4S may be able to do that, but not sure if that
> is in the stock L4S tree or if it requires combining some
> patches.
> 
> BBR can use pacing, but iirc it does not use ECN, which may be
> fixed in bbr3, but I dont know if that has hit a mainstream
> kernel yet.
> 

A reason I added fq-rate-step is to allow network engineers to validate 
source pacing as the best option for e2e link treatment. This combined 
with bytes/packets inflight, cwnd and RTT should allow one to test the 
effectiveness of pacing and then push for ECN or AccurateECN controlled 
pacing, a la L4S. It's an interim test for ECN (including L4S) 
"non-believers" so to speak.

After that, I plan to add CE bit instrumentation. But that requires a 
marking plane (e.g. iproute2 diffs) and a CCA both coming from standard 
kernels. Testing private kernels isn't sufficiently conclusive and 
something I try to avoid if possible.

Note: I added support for TCP_TX_DELAY (thanks to the prescience of Eric 
Dumazet) so one can affect different RTTs per socket. It can be set on 
client, server or both.

--tcp-tx-delay n,[<prob>]
Set TCP_TX_DELAY on the socket. Delay units are milliseconds and 
probability is prob >= 0 and prob <= 1. Values takes float. See Notes 
for qdisc requirements.

(Ignore the probability part for now as it's not fully vetted and 
doesn't really work.)

Bob