XenAPI VCPU information unavailable

[Bob Ball]

Debugging output from the resource tracker
2013-06-19 10:56:53.538 DEBUG nova.compute.resource_tracker [-] Hypervisor: VCPU information unavailable from (pid=4514) _report_hypervisor_resource_view /opt/stack/nova/nova/compute/resource_tracker.py:339

This is because get_available_resource in the XenAPI driver returns 0 for vCPUs present and vCPUs in use.

We should instead have a maximum number of vCPUs - probably a multiple of the pCPUs - and report the number of vCPUs accurately.

At the same time, we should consider removing the xenhost xapi plugin host_data call, and just calling "host.get_rec" directly from nova-compute.

[John Garbutt (johngarbutt)]

Looking at Horizon, this now looks bad. Need to report vCPUs

vCPU info missing, and storage (in use) info looks a bit dodgy too.

[Bob Ball (bob-ball)]

This is medium because it breaks a CPU scheduler filter

[Aditi Raveesh (aditirav)]

From the VM record, it is possible to get the max number of vCPUs and the percentage of each being used.
We can derive the number of vCPUs used, using this information.

Is this the only issue being faced?

[Bob Ball (bob-ball)]

This defect is talking about the host statistics returned by http://git.openstack.org/cgit/openstack/nova/tree/nova/virt/xenapi/driver.py#n436

Yes, you can get per VM statistics - see http://git.openstack.org/cgit/openstack/nova/tree/nova/virt/xenapi/vm_utils.py#n1703. However, if you list_vms then for each ref call "VM.get_VCPUs_max" then that's all that's needed for the total vCPUs used.

I'm not quite sure what should be returned here - looking at the libvirt driver as an example it sets vCPU max to the number of physical processors - which doesn't allow operators to decide how much to over-commit by, and you will easily exceed this "maximum"

[Bob Ball (bob-ball)]

The host CPU utilisation can probably be retrieved from host, getting host_CPUs, then for each of them reading the "utilisation" field" - see http://docs.vmd.citrix.com/XenServer/6.1.0/1.0/en_gb/api/?c=host and http://docs.vmd.citrix.com/XenServer/6.1.0/1.0/en_gb/api/?c=host_cpu

The VM CPU utilisation can likewise be obtained, e.g. xe vm-param-get uuid=82cdbbaf-338f-85d4-82d2-8f4f00a0971d param-name=VCPUs-utilisation. In XenAPI that's obtained from the VM-metrics record referenced from the VM object.

[Bob Ball (bob-ball)]

Sorry - I should have said that given we're primarily interested in the host, I think getting the metrics from the host over CPU utilisation will be closer than adding up all of the vCPU metrics. That'll include any usage in Dom0 such as that used for qemu.

[Aditi Raveesh (aditirav)]

@Bob Ball, yes using the host_cpu information makes sense.

[OpenStack Infra (hudson-openstack)]

Reviewed: https://review.openstack.org/66597
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=8761a7b9c8a967d2b5bf7b690115af98ee0cbd0e
Submitter: Jenkins
Branch: master

commit 8761a7b9c8a967d2b5bf7b690115af98ee0cbd0e
Author: Bob Ball <email address hidden>
Date: Tue Jan 14 15:07:39 2014 +0000

    XenAPI: Report the CPU details correctly

    Driver was not filling in the cpu_count in the correct field.
    Also added VM vcpu count

    Change-Id: I6339d7fd4faffcd5320fd86b55ff3a5e3515e2bf
    Closes-bug: 1192528