[Openstack] Scheduler try and error with compute nodes in different subnets/physnet's

Jay Pipes jaypipes at gmail.com
Tue Sep 16 06:09:32 UTC 2014


On 09/16/2014 02:07 AM, Chris wrote:
> Hello Jay,
> As far as I understand each compute node sends the current available
> resources to the management node.
> This could also include the physnet, so the scheduler ignore the not
> matching compute nodes. Or the compute nodes sends the physnet at the first
> registration in the management node.
> But to be honest I not have the deep insights to make a proper suggestion
> here.

The physnet isn't really a resource though. I think the only thing I can 
think of is perhaps you should use host aggregates and group your 
like-physnetted compute nodes into different host aggregates. The 
problem I still see with that is that there is no way for the scheduler 
to understand which physnet an instance "belongs on"...

-jay

> -----Original Message-----
> From: Jay Pipes [mailto:jaypipes at gmail.com]
> Sent: Tuesday, September 16, 2014 11:54
> To: openstack at lists.openstack.org
> Subject: Re: [Openstack] Scheduler try and error with compute nodes in
> different subnets/physnet's
>
> On 09/15/2014 11:26 PM, Chris wrote:
>> Hello,
>>
>> We have a OpenStack setup with a large number of compute nodes spread
>> in different subnets which are represented as different physnet's in
> OpenStack.
>>
>> When we start an instance we see that the scheduler choose a compute
>> node and tries to spawn the instance, then it sees its not in the
>> right physnet and choose a different compute node.
>
> How would the scheduler know which compute node is "in the right physnet"?
> In other words, when an instance is launched, how does the scheduler know
> what is the "correct physnet" for that type of instance?
>
>> In our case this takes around 4 - 10 attempts. All this attempts
>> counts for the "scheduler_max_attempts" which default value is 3. We
>> increase this value to prevent errors, but it's still very inefficient
>> especially in a large environment.
>>
>> Is there a way that the scheduler knows the physnet/subnet position of
>> the compute nodes before the instance tries to spawn?
>
>
>
> _______________________________________________
> 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
>
>





More information about the Openstack mailing list