sysfs fails to start after setting cpu afinity >32 cpus

[Andrew Woodward]

sysfs fails to start after the cpu affinity is set. When $::processorcount is grater than 32

https://review.openstack.org/#/c/184312/ added sysfs. but in some cases/hardware its set to 'ffffffffff' which causes sysfs to fail to start and the deployment fails.

The value for cpu_affinity' /var/log/puppet.log
2015-06-02 20:26:54 +0000 /Stage[main]/Main/Notify[The value for cpu_affinity_hex is ffffffffff] (info): Starting to evaluate the resource
2015-06-02 20:26:54 +0000 Puppet (notice): The value for cpu_affinity_hex is ffffffffff
2015-06-02 20:26:54 +0000 /Stage[main]/Main/Notify[The value for cpu_affinity_hex is ffffffffff]/message (notice): defined 'message' as 'The value for cpu_affinity_hex is ffffffffff'
2015-06-02 20:26:54 +0000 /Stage[main]/Main/Notify[The value for cpu_affinity_hex is ffffffffff] (debug): The container Class[Main] will propagate my refresh event
2015-06-02 20:26:54 +0000 /Stage[main]/Main/Notify[The value for cpu_affinity_hex is ffffffffff] (info): Evaluated in 0.00 seconds
2015-06-02 20:27:15 +0000 /Stage[main]/Main/Notify[The value for cpu_affinity_hex is ffffffffff] (info): Starting to evaluate the resource
2015-06-02 20:27:15 +0000 Puppet (notice): The value for cpu_affinity_hex is ffffffffff

[Dmitry Ilyin (idv1985)]

ffffffffff means 40 cpus! Are there really 40 cpus?

[Dmitry Ilyin (idv1985)]

perhaps hyper threading or something. Maybe it would be better to use physicalprocessorcount instead of processor count

[Dmitry Ilyin (idv1985)]

And we should also exclude net interfaces i guess

[Dmitry Ilyin (idv1985)]

vnet interfaces created by compute's vas

[Eric Jacobs (ejacobs)]

Yes, there are really 40 cpus in /proc/cpuinfo. I'm not sure if it's hyperthreading or physical. I've seen boxes with 64 physical cores before. To get around this, I just discard all cpus >32 (for now).

[Alex Schultz (alex-schultz)]

Review opened, https://review.openstack.org/#/c/187801/

[Bogdan Dobrelya (bogdando)]

The 6.0.x looks like a good candidate for the backport, see https://bugs.launchpad.net/fuel/+bug/1401925

[Bogdan Dobrelya (bogdando)]

Oh, my wrong - won't fix for the 6.0.x as related bug is actually https://bugs.launchpad.net/fuel/+bug/1456587 which is also won't fix for the 6.0.x

[OpenStack Infra (hudson-openstack)]

Reviewed: https://review.openstack.org/187801
Committed: https://git.openstack.org/cgit/stackforge/fuel-library/commit/?id=87f429928fe6e91952698a34fa356bbb6918eee0
Submitter: Jenkins
Branch: master

commit 87f429928fe6e91952698a34fa356bbb6918eee0
Author: Eric Jacobs <email address hidden>
Date: Tue Jun 2 18:40:36 2015 -0400

    Don't use more than 32 cpus for network queue affinities

    IF we have more than 32 cpus, discard any cpus > 32 for network queue
    cpu affinities.

    Closes-Bug: 1461306
    Change-Id: I6aff2af39f25a32c710d5cdf9708dad6b2a201c1

[Andrew Woodward (xarses)]

sysfs wasn't added to 6.0 so it's invalid for it, not won't fix