[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <bcd3b198697e16059ec69566251ad23c4c78e7a7.camel@trillion01.com>
Date: Tue, 30 Jul 2024 19:14:03 -0400
From: Olivier Langlois <olivier@...llion01.com>
To: Pavel Begunkov <asml.silence@...il.com>, io-uring@...r.kernel.org
Cc: netdev@...r.kernel.org
Subject: Re: io_uring NAPI busy poll RCU is causing 50 context
switches/second to my sqpoll thread
On Tue, 2024-07-30 at 21:25 +0100, Pavel Begunkov wrote:
>
> Removing an entry or two once every minute is definitely not
> going to take 50% CPU, RCU machinery is running in background
> regardless of whether io_uring uses it or not, and it's pretty
> cheap considering ammortisation.
>
> If anything it more sounds from your explanation like the
> scheduler makes a wrong decision and schedules out the sqpoll
> thread even though it could continue to run, but that's need
> a confirmation. Does the CPU your SQPOLL is pinned to stays
> 100% utilised?
Here are the facts as they are documented in the github issue.
1. despite thinking that I was doing NAPI busy polling, I was not
because my ring was not receiving any sqe after its initial setup.
This is what the patch developped with your input
https://lore.kernel.org/io-uring/382791dc97d208d88ee31e5ebb5b661a0453fb79.1722374371.git.olivier@trillion01.com/T/#u
is addressing
(BTW, I should check if there is such a thing, but I would love to know
if the net code is exposing a tracepoint when napi_busy_poll is called
because it is very tricky to know if it is done for real or not)
2. the moment a second ring has been attached to the sqpoll thread that
was receving a lot of sqe, the NAPI busy loop started to be made for
real and the sqpoll cpu usage unexplicably dropped from 99% to 55%
3. here is my kernel cmdline:
hugepages=72 isolcpus=0,1,2 nohz_full=0,1,2 rcu_nocbs=0,1,2
rcu_nocb_poll irqaffinity=3 idle=nomwait processor.max_cstate=1
intel_idle.max_cstate=1 nmi_watchdog=0
there is absolutely nothing else on CPU0 where the sqpoll thread
affinity is set to run.
4. I got the idea of doing this:
echo common_pid == sqpoll_pid > /sys/kernel/tracing/events/sched/filter
echo 1 > /sys/kernel/tracing/events/sched/sched_switch/enable
and I have recorded over 1,000 context switches in 23 seconds with RCU
related kernel threads.
5. just for the fun of checking out, I have disabled NAPI polling on my
io_uring rings and the sqpoll thread magically returned to 99% CPU
usage from 55%...
I am open to other explanations for what I have observed but my current
conclusion is based on what I am able to see... the evidence appears
very convincing to me...
> >
> > So is this a good idea in your opinion?
>
> I believe that's a good thing, I've been prototyping a similar
> if not the same approach just today, i.e. user [un]registers
> napi instance by id you can get with SO_INCOMING_NAPI_ID.
>
this is fantastic!
I am super happy to see all this NAPI busy polling feature interest and
activity which is a feature that I am very fond with (along with
io_uring)
I am looking forward collaborating with you Pavel to make io_uring the
best NAPI busy polling goto solution!
Greetings,
Olivier
Powered by blists - more mailing lists