Re: Low TCP throughput due to vmpressure with swap enabled

[Shakeel Butt <shakeelb-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org>]

On Mon, Nov 28, 2022 at 01:07:25PM -0500, Johannes Weiner wrote:
> 
[...]
> > With the patch applied I'm capped at ~120MB/s, which is a symptom of a
> > clamped window.
> > 
> > I can't find any sockets with memcg->socket_pressure = 1, but at the
> > same time I only see the following rcv_ssthresh assigned to sockets:
> 
> Hm, I don't see how socket accounting would alter the network behavior
> other than through socket_pressure=1.
> 

I think what is happening is that the tcp stack is calling
tcp_under_memory_pressure() and making decisions without going through
the memcg charge codepath which set or reset memcg->socket_pressure.
Most probably the socket is clamped due to memcg->socket_pressure and
then the kernel never tried to grow its buffers because
memcg->socket_pressure is still set and thus never tried the memcg
charge codepath which would have reset memcg->socket_pressure. (Maybe)
That is my guess but network experts CCed can correct me.

Anyways, I don't think the pressure mechanism which relies on successful
charging will work. I am brainstorming towards memory.high based network
throttling. Basically use penalty_jiffies (or something similar) to set
memcg->socket_pressure. However I want this to be opt-in as we do have
applications which prefer to be killed than be throttled. So, still
working on the fine details how this can be done without introducing a
rigid API.