Re: [PATCH v3 00/10] tcp: Fix throughput issues with non-local peers

[Stefano Brivio <sbrivio@redhat.com>]

On Mon, 08 Dec 2025 01:51:48 -0700
Max Chernoff <git@maxchernoff.ca> wrote:

> Hi Stefano,
> 
> On Mon, 2025-12-08 at 09:25 +0100, Stefano Brivio wrote:
> > but, in general, I don't think things can work reliably
> > with the values you shared for tcp_notsent_lowat.  
> 
> Ok, that works for me. I know very little about TCP, so I just blindly
> copied that value for tcp_notsent_lowat from
> 
>     https://blog.cloudflare.com/http-2-prioritization-with-nginx/

Right, I guess you mentioned it already. I still need to find some time
to actually look into it in better detail and understand what's still
true now (note that that article is from 2018 and there have been very
relevant kernel changes against "bufferbloat" in between). At that
point:

> but if that's incompatible with pasta, then I have no problem resetting
> tcp_notsent_lowat back to the kernel default.
> 
> A random web search makes it look like changing tcp_notsent_lowat is
> somewhat common
> 
>     https://www.google.com/search?q=tcp_notsent_lowat%3D131072
> 
>     https://github.com/search?q=tcp_notsent_lowat%3D131072+NOT+is%3Afork&type=code
> 
> so maybe it would be a good idea for pasta to either use setsockopt to
> override it, or to print a warning on startup if the sysctl is set too
> low?

...I would also get an opinion on this. I should be able to get back to
you in a few days.

> > Does this (upload now taking longer/timing out with 50 ms RTT) also
> > happen without "custom" values for tcp_notsent_lowat?
> >
> > I tested things quite extensively in that RTT region (without custom
> > sysctl values) and the improvement looks rather consistent to me.  
> 
> Ok, with tcp_notsent_lowat reset to the Fedora defaults, the upload
> speeds with large RTTs do indeed look *much* better

Ok, great, thanks for confirming. That's all I'm trying to fix for the
moment. :)

-- 
Stefano