Linaro Linux

Processes do not distribute across cores when running CoreMark

Bug #754248 reported by Michael Hope
18
This bug affects 1 person
Affects Status Importance Assigned to Milestone
Linaro Linux
Triaged
Low
Amit Kucheria

Bug Description

Sorry for the poor report but I've run out of time to track this down.

I've seen this a couple of times now when running the ti-ubuntu-2.6.35-ti903.13+release5-445-g63213d4 Ubuntu kernel. I believe I've seen it on Linux Linaro 2.6.38 as well but aren't sure.

CoreMark is a synthetic benchmark used for exercising the toolchain and processor. It can run in a pthread or fork() based mode on SMP machines to exercise the different cores. Some times the system gets into a state where the second core on a PandaBoard does not spin up and the threads or processes stay on the one core. This halves the benchmark results.

To download CoreMark see:
 http://coremark.org/home.php

To build it for a SMP system try make XCFLAGS="-DMULTITHREAD=2 -DUSE_PTHREAD"

The system only gets into this state now and again. When it does, two runs out of three show the fault. I'm currently running the attached .config with PM and the governor turned off to see if the problem goes away.

Revision history for this message
Michael Hope (michaelh1) wrote :
Revision history for this message
warmcat (andy-warmcat) wrote :

If you have CPU_FREQ running that seems to be kind of broken at the moment.

/sys/devices/system/cpu/cpu0
/sys/devices/system/cpu/cpu0/cpufreq
/sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_min_freq
/sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_max_freq
/sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_transition_latency
/sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_cur_freq
/sys/devices/system/cpu/cpu1
/sys/devices/system/cpu/cpu1/cpufreq

It doesn't seem to attach to cpu1 like it does on a laptop for example.

Revision history for this message
Amit Kucheria (amitk) wrote :

What kernel is this? Sounds like a dodgy version....

Also, is this on Panda? Could you update the subject to reflect that.

Revision history for this message
Michael Hope (michaelh1) wrote :

I added [Panda] to the title. As in the subject, the kernel I saw this in is ti-ubuntu-2.6.35-ti903.13+release5-445-g63213d4 from http://dev.omapzoom.org/. I believe I have also seen this on Linaro 2.6.38 but currently can't confirm it.

Revision history for this message
Amit Kucheria (amitk) wrote : Re: [Bug 754248] Re: Processes do not distribute across cores when running CoreMark

On 11 Apr 14, Michael Hope wrote:
> I added [Panda] to the title. As in the subject, the kernel I saw this
> in is ti-ubuntu-2.6.35-ti903.13+release5-445-g63213d4 from
> http://dev.omapzoom.org/. I believe I have also seen this on Linaro
> 2.6.38 but currently can't confirm it.

Running the omapzoom kernel is akin to running the latest crack of the day
(with all due respect to TI developers). Is there any specific feature in
there you're looking for?

In any case, cpufreq is WIP on Panda. So I'm not completely surprised. This
bug would need some detailed analysis from someone that understands that
code. Vishwa might be that someone. Adding him.

Revision history for this message
warmcat (andy-warmcat) wrote :

@Amit: this is the behaviour on linux-linaro-2.6.38 since npitre took the (now rejected upstream) cpufreq patches.

@Michael: it makes sense to use 2.6.38 kernel basis on Panda now I think. You can try npitre's linux-linaro-2.6.38 and disable CPU_FREQ in config

Ubuntu themselves are now sourcing Panda kernel from my tree

http://git.linaro.org/gitweb?p=people/andygreen/kernel-tilt.git;a=shortlog;h=refs/heads/for-ubuntu

which is based on linux-linaro-2.6.38. Any bugfixes are going to be happening on these trees I think.

Linus Walleij (triad)
Changed in linux-linaro:
assignee: nobody → Amit Kucheria (amitk)
Jakub Pavelek (jakub-pavelek)
Changed in linux-linaro:
status: New → Triaged
Jakub Pavelek (jakub-pavelek)
Changed in linux-linaro:
importance: Undecided → Low
To post a comment you must log in.
This report contains Public information  
Everyone can see this information.
Subscribing...

Other bug subscribers

Bug attachments

Remote bug watches

Bug watches keep track of this bug in other bug trackers.