[Openstack] [OpenStack][Nova] NUMA Topology together with requested PCI devices Schedule Failed

Chris Friesen chris.friesen at windriver.com
Mon Mar 13 14:26:29 UTC 2017


On 03/13/2017 03:06 AM, Xu, Rongjie (Nokia - CN/Hangzhou) wrote:
> Hi,
> I want to know whether SRIOV VF can be used cross socket (e.g PF is on socket 0,
> so its VFs also on socket 0, and VM is on socket 1).
> My testing is to launch several VMs with dedicated vCPU with VF. Firstly, I saw
> VMs are located on socket 0. When the CPUs on socket 0 are used up, I expect VM
> can be scheduled to socket 1 or some error happens. In indeed, I got following
> errors:
> Insufficient compute resources: Requested instance NUMA topology together with
> requested PCI devices cannot fit the g", "code": 500, "details": "  File
> \"/usr/lib/python2.7/dist-packages/nova/conductor/manager.py\", line 393, in
> build_instances
> But what I cannot understand is what does “NUMA topology” indicate here? How it
> is related to NUMA topology? Can anybody help explain?

Currently nova does not support cross-socket usage, and the error message you're 
seeing reflects that.  (Asking for pci devices results in specifying some 
restrictions on the guest, hence the "numa" related error message.)

There's a proposed spec to allow this requirement to be relaxed, see

https://review.openstack.org/#/c/361140/18/specs/pike/approved/share-pci-between-numa-nodes.rst

Chris




More information about the Openstack mailing list