Added section about NUMA

Adds some conceptual information about NUMA, and discusses how nova
considers NUMA topology when booting instances. (This is a starting point,
I'm happy for people to move this to a different section, or change wording).

Change-Id: I328831fb3aef9531ee924aa86726277b7eac0cb5
Closes-Bug: 1369207
This commit is contained in:
Lana Brindley 2015-09-03 15:09:54 +10:00
parent d6d10a5fb7
commit 9a2a1a375d
2 changed files with 26 additions and 0 deletions

View File

@ -8,6 +8,7 @@ This section includes tasks specific to the OpenStack environment.
:maxdepth: 2
cli_nova_specify_host.rst
cli_nova_numa_libvirt.rst
cli_nova_evacuate.rst
cli_nova_migrate.rst
cli_admin_manage_ip_addresses.rst

View File

@ -0,0 +1,25 @@
=============================================
Consider NUMA topology when booting instances
=============================================
NUMA topology can exist on both the physical hardware of the host, and the
virtual hardware of the instance. OpenStack Compute uses libvirt to tune
instances to take advantage of NUMA topologies. The libvirt driver boot
process looks at the NUMA topology field of both the instance and the host it
is being booted on, and uses that information to generate an appropriate
configuration.
If the host is NUMA capable, but the instance has not requested a NUMA
topology, Compute attempts to pack the instance into a single cell.
If this fails, though, Compute will not continue to try.
If the host is NUMA capable, and the instance has requested a specific NUMA
topology, Compute will try to pin the vCPUs of different NUMA cells
on the instance to the corresponding NUMA cells on the host. It will also
expose the NUMA topology of the instance to the guest OS.
If you want Compute to pin a particular vCPU as part of this process,
set the ``vcpu_pin_set`` parameter in the :file:`nova.conf` configuration
file. For more information about the ``vcpu_pin_set`` parameter, see the
Configuration Reference Guide.