Re: Regression with idle cpu cycle handling in 2.6.24 (compared to 2.6.22)

["dAniel hAhler" <dhahler@googlemail.com>]

Hello,

I've now found the reason and a workaround for this. Apparently, it's
related to CONFIG_FAIR_USER_SCHED and can be worked around by
assigning a really small value to the boinc users cpu_share (125 is
the uid of "boinc"):
$ echo 2 | sudo tee /sys/kernel/uids/125/cpu_share

While looking around, I've found the following patch, which seems to
address this:
http://www.ussg.iu.edu/hypermail/linux/kernel/0710.3/3849.html

It has been posted here, but without any response.

btw: writing 1 into "cpu_share" totally locks up the computer!


Cheers.

On Jan 19, 2008 3:05 PM, I wrote:
> Hello,
>
> I have BOINC running in the background with niceness 19.
> With a 2.6.22 kernel, only idle cpu cycles get assigned to this process, as
> expected.
>
> But with the 2.6.24 kernel, the BOINC process gets at least about half of
> all CPU cycles, even if there's another process (owned by another user)
> requesting CPU cycles (e.g. "cat /dev/urandom > /dev/null")
>
> This happens with the Ubuntu kernel (from Hardy) and the daily builds from
> http://kernel-archive.buildserver.net/debian-kernel/ (where I've just tested
> rc8 now).
>
> It appears that every user (here "boinc" and my user) get the same portion
> of the overall CPU cycles, regardless of the process niceness.
>
> Is this expected behaviour?
>
> I'm using an AMD64 3000+ processor. Please ask for additional information, if
> you need it (e.g. kernel config).
>
>
> TESTCASE:
> $ cat /dev/urandom > /dev/null &
> $ sudo -u another_user nice -n 19 python -c 'i = 0;
> while 1:
>   i += i
> '
>
> EXPECTED RESULT:
> The niced process should get nearly no CPU cycles.
>
> ACTUAL RESULT:
> The niced process gets about half of the CPU cycles (according to "top").
>
>
> The bug has been reported for Ubuntu on https://launchpad.net/bugs/177713
>
>
> --
> http://daniel.hahler.de/
>



-- 
http://daniel.hahler.de/