Nova libvirt driver does not bind guest memory to NUMA nodes

[Daniel Berrange]

After testing the libvirt NUMA support I noticed that the guest XML is missing rules to explicitly bind guest NUMA nodes to host NUMA nodes.

The vCPUs are still all bound to host NUMA nodes, so the kernel will prefer allocation from those nodes, but it is a much better to be explicit about the binding to prevent the kernel from falling back to allocate memory from a non-local node, as that would confuse Nova's tracking of NUMA allocation.

[OpenStack Infra (hudson-openstack)]

Related fix proposed to branch: master
Review: https://review.openstack.org/130820

[OpenStack Infra (hudson-openstack)]

Fix proposed to branch: master
Review: https://review.openstack.org/131131

[OpenStack Infra (hudson-openstack)]

Reviewed: https://review.openstack.org/130820
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=943f02f1a62f7eb131f62729d83d97ad431ee89f
Submitter: Jenkins
Branch: master

commit 943f02f1a62f7eb131f62729d83d97ad431ee89f
Author: Daniel P. Berrange <email address hidden>
Date: Fri Oct 24 16:12:58 2014 +0100

    libvirt: add classes for NUMA memory binding configuration

    To allow specification of strict memory binding for NUMA
    nodes, add support for libvirt's <numatune> XML element.

    Related-bug: #1385308
    Change-Id: I08f01a50b92ca648b803f94ef028bc2f10e6f6a8

[OpenStack Infra (hudson-openstack)]

Reviewed: https://review.openstack.org/131131
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=d9276355d9468d62a4c3150bf0401ee0f3ccd031
Submitter: Jenkins
Branch: master

commit d9276355d9468d62a4c3150bf0401ee0f3ccd031
Author: Daniel P. Berrange <email address hidden>
Date: Mon Oct 27 11:30:24 2014 +0000

    libvirt: set NUMA memory allocation policy for instances

    When NUMA placement is in effect, Nova configures KVM guest vCPUs
    to be placed on certain host NUMA nodes. It also provides guest
    NUMA topology of vCPUs and RAM. The kernel will normally satisfy
    memory allocations from the host NUMA node that the guest vCPU
    is running on, but this is only a preference. Nova needs to set
    a strict policy to guarantee guest RAM is always allocated from
    the desired host NUMA node. This change ensures that by setting
    the XML

      <numatune>
        <memory mode='strict' nodeset='0-1'/>
        <memnode cellid='0' mode='strict' nodeset='0'/>
        <memnode cellid='1' mode='strict' nodeset='1'/>
      </numatune>

    Closes-bug: #1385308
    Change-Id: I02c301ae648235dc48527559a5d3c3d45de859ac