Re: SLUB: purpose of sysfs events on cache creation/removal

[Michal Koutný]

Hello.

On Thu, Jan 09, 2020 at 11:44:15AM -0800, Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> I looked at it - there wasn't really any compelling followup.
FTR, I noticed udevd consuming non-negligible CPU cycles when doing some
cgroup stress testing. And even extrapolating to less artificial
situations, the udev events seem to cause useless tickling of udevd.

I used the simple script below
cat measure.sh <<EOD
sample() {
        local n=$(echo|awk "END {print int(40/$1)}")

        for i in $(seq $n) ; do
                mkdir /sys/fs/cgroup/memory/grp1 ;
                echo 0 >/sys/fs/cgroup/memory/grp1/cgroup.procs ;
                /usr/bin/sleep $1 ;
                echo 0 >/sys/fs/cgroup/memory/cgroup.procs ;
                rmdir /sys/fs/cgroup/memory/grp1 ;
        done
}

for d in 0.004 0.008 0.016 0.032 0.064 0.128 0.256 0.5 1 ; do
        echo 0 >/sys/fs/cgroup/cpuacct/system.slice/systemd-udevd.service/cpuacct.usage
        time sample $d 2>&1 | grep real
        echo -n "udev "
        cat /sys/fs/cgroup/cpuacct/system.slice/systemd-udevd.service/cpuacct.usage
done
EOD

and I drew the following ballpark conclusion:
1.7% CPU time at 1 event/s -> 60 event/s 100% cpu

(The event is one mkdir/migrate/rmdir sequence. Numbers are from dummy
test VM, so take with a grain of salt.)


> If this change should be pursued then can we please have a formal
> resend?
Who's supposed to do that?

Regards,
Michal

Attachment: signature.asc
Description: Digital signature