(In reply to VIKRANT from comment #0)
> Description of problem:
> 
> Not able to set the huge pages in guest OS if the cpu_model parameter in set
> on compute node. If this parameter is not set on compute node then able to
> set the huge pages in guest OS without any issue. 
> 
> 
> 
> Version-Release number of selected component (if applicable):
> RHEL OSP 7
> 
> Other versions are likely impacted due to this issue.
> 
> How reproducible:
> Everytime. 
> 
> 
> Steps to Reproduce:
> 1. Set the cpu_model" parameter in /etc/nova/nova.conf file of compute node.
> 2. Try to set huge pages on guest instance but it's not getting set.
> 3. Remove the cpu_model parameter now we are able to set the huge pages in 
> guest OS without any issue.
> 
> Actual results:
> We are not able to set guest OS huge pages until cpu_model is set on compute
> node. 
> 
> Expected results:
> We should be able to set guest OS huge pages irrespective of cpu_model is
> set on compute node or not. 

It isn't quite that simple. If using 2 MB huge pages, you can use *any* CPU model as no special features are required for that to work. If you want to use 1 GB huge pages, then uou need to pick a CPU model that contains the "pdpe1gb" flag.  *none* of the Intel CPU models supported by QEMU include that flag. So realistically at this time you need to use the "host-passthrough" feature.

Nova *does* need a feature enhancement to extend the "cpu_model" syntax to allow us to request extra CPU flags beyond what the basic model allows. eg so you could do

cpu_model=Haswell
cpu_model_extra_flags="pdpe1gb"

but that will obviously need some work upstream to support

Daniel,

Thanks. When you saying " *none* of the Intel CPU models supported by QEMU include that flag." does that mean this limitation is present outside nova ? But if the Cu is not setting "cpu_model" then in that case huge pages are working fine.

(In reply to VIKRANT from comment #3)
> Daniel,
> 
> Thanks. When you saying " *none* of the Intel CPU models supported by QEMU
> include that flag." does that mean this limitation is present outside nova ?
> But if the Cu is not setting "cpu_model" then in that case huge pages are
> working fine.

The features included in the named CPU models are defined by QEMU, not Nova, so apply to any app using those models. Only the Opteron_G4 and Opteron_G5 CPU models include the pdpe1gb feature flag. For every other CPU model you need to request the flag manually, which nova does not yet support

For Nova, if neither 'cpu_mode' or 'cpu_model' are set in nova.conf, then Nova will fallback to use "host-model" which will include the flag.

I'm going to move this one to documentation for now, as while there is a long term need to provide some more flexibility in Nova here it doesn't necessarily seem imminent.

Cloned this back to Nova to track future CPU flag enhancement, since this is not the first time this general requirement has come up

 https://bugzilla.redhat.com/show_bug.cgi?id=1398343

(In reply to Stephen Gordon from comment #5)
> I'm going to move this one to documentation for now, as while there is a
> long term need to provide some more flexibility in Nova here it doesn't
> necessarily seem imminent.

Steve, are there specific updates you're looking to have added to the documentation to cover this bug?

I'll update the summary accordingly.

Basically the documentation around the use of huge pages, specifically 1G ones, needs to call out the constraints Dan mentions in comment # 2.

At this time this effectively means they must use libvirt_cpu_mode=host-passthrough on the compute nodes if the intent is to use 1G huge pages in the guests, this has ramifications for cluster compatibility for live migration (see also: https://wiki.openstack.org/wiki/LibvirtXMLCPUModel).

Thanks, Steve. Updating the summary to reflect the requirement.

SME: Kashyap Chamarty <kchamart>

Updated content available on the Customer Portal:

https://access.redhat.com/documentation/en-us/red_hat_openstack_platform/16.0/html-single/instances_and_images_guide/index#ch-huge-pages