[Openstack] [OpenStack] [nova] How can I release PCI device resources without delete instance?

Eddie Yen missile0407 at gmail.com
Thu Jul 6 06:17:22 UTC 2017


Hi, now I got another problem.

I have two models of GPU devices and I set both in pci_alias and
pcli_passthrogh_whitelist on controller and compute node(with this two
GPUs).
Now I removed one of GPU and delete its data in nova.conf, then restart
nova-api, nova-scheduler (Controller) and nova-compute(Compute)

But when I check MySQL, the GPU info which I already removed still in
pci_devices table.
I remember there's a bug about this case but it already fixed.

How can I fix this issue?

Thanks,


2017-07-06 11:42 GMT+08:00 Eddie Yen <missile0407 at gmail.com>:

> Thanks Kevin,
>
> This is another good option to solve IP problem for me. I'll note this.
>
> Eddie.
>
> 2017-07-06 11:10 GMT+08:00 Kevin Benton <kevin at benton.pub>:
>
>> As a stop gap for the IP problem, keep in mind that you can create a
>> neutron port separately and boot the instance to it. Then when you delete
>> the instance the port will remain and hold onto the IP so you can boot to
>> it with a different instance later.
>>
>> On Jul 5, 2017 18:31, "Eddie Yen" <missile0407 at gmail.com> wrote:
>>
>>> Hi, thanks for reply first!
>>>
>>> After I tried shelve instance, yes it really release GPU resources.
>>> But as Dinesh said, can't unshelve this instance if all GPUs occupied.
>>>
>>> Before known about shelve instance, I have do few approach.
>>>
>>> One is using "resize" the instance's flavor to non-GPU flavor, but can't
>>> do this if all GPUs occupied.
>>>
>>> Another one is like Jay's recommend is that create instance as volume,
>>> and delete instance but keep volume if this VM not going to use GPU.
>>> Then re-create instance with previous volume and non-GPU flavor.
>>> And yes, this approach have IP address problem that may cause
>>> application atypical.
>>>
>>> I'll try how to make my application not to rely on a persistent IP.
>>>
>>> Thanks,
>>> Eddie.
>>>
>>> 2017-07-05 19:36 GMT+08:00 Jay Pipes <jaypipes at gmail.com>:
>>>
>>>> On 07/05/2017 04:18 AM, Eddie Yen wrote:
>>>>
>>>>> Hi everyone,
>>>>>
>>>>> I'm using OpenStack Mitaka (which deployed from Fuel 9.2) and doing
>>>>> GPU things.
>>>>>
>>>>> I got a problem is I need to delete current instance which contains
>>>>> GPU to release device if I want assign GPU to another new instance
>>>>> temperately.
>>>>>
>>>>> I'll got "No valid host was found" if I creating new instance contains
>>>>> GPU flavor without delete present instance, even the instance is shutdown.
>>>>>
>>>>> Is there any way that I just shutdown the instance rather than delete
>>>>> it to release GPU device?
>>>>>
>>>>
>>>> As Dinesh mentioned you *can* use shelve for this, but frankly, I think
>>>> the shelve API leads to more problems than it solves (see his response
>>>> about needing to delete the new instance before unshelving).
>>>>
>>>> I'd recommend redesigning your application to be more cloud-native. In
>>>> other words, separate operating system state from application state, use
>>>> volumes for all persistent application state, and do not rely on a
>>>> persistent IP address. [1]
>>>>
>>>> Once you've done that, you can just treat your VMs like cattle and
>>>> terminate them.
>>>>
>>>> Best,
>>>> -jay
>>>>
>>>> [1] Please note I did not use the word "container" in this description
>>>> of cloud-native application.
>>>>
>>>> _______________________________________________
>>>> Mailing list: http://lists.openstack.org/cgi
>>>> -bin/mailman/listinfo/openstack
>>>> Post to     : openstack at lists.openstack.org
>>>> Unsubscribe : http://lists.openstack.org/cgi
>>>> -bin/mailman/listinfo/openstack
>>>>
>>>
>>>
>>> _______________________________________________
>>> Mailing list: http://lists.openstack.org/cgi
>>> -bin/mailman/listinfo/openstack
>>> Post to     : openstack at lists.openstack.org
>>> Unsubscribe : http://lists.openstack.org/cgi
>>> -bin/mailman/listinfo/openstack
>>>
>>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack/attachments/20170706/6269c8a4/attachment.html>


More information about the Openstack mailing list