Re: [RFC PATCH v4 3/3] softirq: defer softirq processing to ksoftirqd if CPU is busy with RT

[Qais Yousef <qyousef@layalina.io>]

(note the change in email)

On 10/17/22 17:04, John Stultz wrote:
> On Mon, Oct 17, 2022 at 7:45 AM Qais Yousef <qais.yousef@arm.com> wrote:
> > This time I paid attention to the average as the best case number for vanilla
> > kernel is better:
> >
> >                    |       vanilla      | with softirq patches v4  |
> > -------------------|--------------------|--------------------------|
> >                    |  #1  |  #2  |  #3  |   #1   |   #2   |   #3   |
> > -------------------|------|------|------|--------|--------|--------|
> > t0 avg delay (us)  |31.59 |22.94 |26.50 | 31.81  | 33.57  | 34.90  |
> > t1 avg delay (us)  |16.85 |16.32 |37.16 | 29.05  | 30.51  | 31.65  |
> > t2 avg delay (us)  |25.34 |32.12 |17.40 | 26.76  | 28.28  | 28.56  |
> >
> > It shows that we largely hover around 30us with the patches compared to 16-26us
> > being more prevalent for vanilla kernels.
> >
> > I am not sure I can draw a concrete conclusion from these numbers. It seems
> > I need to run longer than 4 hours to hit the worst case scenario every run on
> > the vanilla kernel. There's an indication that the worst case scenario is
> > harder to hit, and it looks there's a hit on the average delay.
> 
> Thanks so much for running these tests and capturing these detailed numbers!
> 
> I'll have to look further into the average case going up here.
> 
> > I'm losing access to this system from today. I think I'll wait for more
> > feedback on this RFC; and do another round of testing for longer periods of
> > time once there's clearer sense this is indeed the direction we'll be going
> > for.
> 
> Do you mind sending me the script you used to run the test, and I'll
> try to reproduce on some x86 hardware locally?

I ran that in a personal CI setup. I basically do the following 3 in parallel
'scripts':

cyclictest.sh [1]:

	cyclictest -t 3 -p 99 -D 3600 -i 1000 --json=cyclictest.json

iperf.sh [2]:

	iperf -s -D
	iperf -c localhost -u -b 10g -t 3600 -i 1 -P 3

dd.sh [3]:

	while true
	do
		cyclictest_running=`ps -e | grep cyclictest || true`
		if [ "x$cyclictest_running" == "x" ]; then
			break
		fi

		#
		# Run dd
		#
		file="/tmp/myci.dd.file"
		for i in $(seq 3)
		do
			dd if=/dev/zero  of=$file.$i bs=1M count=2048 &
		done
		wait
		rm -f $file*

		sleep 3
	done

[1] https://github.com/qais-yousef/myci-sched-tests/blob/dev/vars/run_cyclictest.groovy
[2] https://github.com/qais-yousef/myci-sched-tests/blob/dev/vars/run_iperf_parallel.groovy
[3] https://github.com/qais-yousef/myci-sched-tests/blob/dev/vars/run_dd_parallel.groovy


Cheers

--
Qais Yousef