Open Stack

On 11 July 2013 02:39, Russell Bryant <rbryant at redhat.com> wrote:

>> We'll probably need something like this for Ironic with persistent
>> volumes on machines - yes its a rare case, but when it matters, it
>> matters a great deal.
>
> I believe you, but I guess I'd like to better understand how this works
> to make sure what gets added actually solves your use case.  Is there
> already support for Cinder managed persistent volumes that live on
> baremetal nodes?

There isn't, but we discussed it with Cinder folk in Portland.

Basic intent is this:
 - we have a cinder 'Ironic' backend.
 - volume requests for the Ironic backend are lazy provisioned: they
just allocate a UUID on creation
 - nova-bm/Ironic will store a volume <-> node mapping
 - 'nova boot' without a volume spec will only select nodes with no
volumes associated to them.
 - 'nova boot' with a volume spec will find an existing node with
those volumes mapped to it, or if none exists create the volume
mapping just-in-time
 - the deployment ramdisk would setup the volume on the hardware
[using hardware RAID]
   - where there isn't hardware RAID we'd let the instance take care
of how to setup the persistent storage - because we don't have a
translation layer in place we can't assume lvm or windows volume
manager or veritas or.....

The obvious gap between intent and implementation here is that choices
about nodes happen in the nova scheduler, so we need the scheduler to
be able to honour four cases:
 - baremetal flavor with no volumes requested, gets a baremetal node
with no volumes mapped
 - baremetal flavor with volumes requested, just one baremetal node
with any of those volumes exist -> that node
 - baremetal flavor with volumes requested, > one baremetal node with
any of those volumes exist -> error
 - baremetal flavor with volumes requested, no nodes with any of those
volumes -> pick any node that has enough disks to supply the volume
definitions

Writing this it seems like the nova scheduler may be a tricky fit;
perhaps we should - again- reevaluate just how this all glues
together?

-Rob
-- 
Robert Collins <rbtcollins at hp.com>
Distinguished Technologist
HP Cloud Services