On Wed, May 31, 2017 at 10:01 PM, Jay Pipes <[email protected]> wrote:
> On 05/31/2017 01:31 AM, Zhenguo Niu wrote: > >> On Wed, May 31, 2017 at 12:20 PM, Ed Leafe <[email protected] <mailto: >> [email protected]>> wrote: >> >> On May 30, 2017, at 9:36 PM, Zhenguo Niu <[email protected] >> <mailto:[email protected]>> wrote: >> >> > as placement is not splitted out from nova now, and there would be >> users who only want a baremetal cloud, so we don't add resources to >> placement yet, but it's easy for us to turn to placement to match the node >> type with mogan flavors. >> >> Placement is a separate service, independent of Nova. It tracks >> Ironic nodes as individual resources, not as a "pretend" VM. The >> Nova integration for selecting an Ironic node as a resource is still >> being developed, as we need to update our view of the mess that is >> "flavors", but the goal is to have a single flavor for each Ironic >> machine type, rather than the current state of flavors pretending >> that an Ironic node is a VM with certain RAM/CPU/disk quantities. >> >> >> Yes, I understand the current efforts of improving the baremetal nodes >> scheduling. It's not conflict with mogan's goal, and when it is done, we >> can share the same scheduling strategy with placement :) >> >> Mogan is a service for a specific group of users who really want a >> baremetal resource instead of a generic compute resource, on API side, we >> can expose RAID, advanced partitions, nics bonding, firmware management, >> and other baremetal specific capabilities to users. And unlike nova's host >> based availability zone, host aggregates, server groups (ironic nodes share >> the same host), mogan can make it possible to divide baremetal nodes into >> such groups, and make Rack aware for affinity and anti-affinity when >> scheduling. >> > Zhenguo Niu brings up a very good point here. Currently, all Ironic nodes > are associated with a single host aggregate in Nova, because of the > vestigial notion that a compute *service* (ala the nova-compute worker) was > equal to the compute *node*. > > In the placement API, of course, there's no such coupling. A placement > aggregate != a Nova host aggregate. > > So, technically Ironic (or Mogan) can call the placement service to create > aggregates that match *its* definition of what an aggregate is (rack, row, > cage, zone, DC, whatever). Furthermore, Ironic (or Mogan) can associate > Ironic baremetal nodes to one or more of those placement aggregates to get > around Nova host aggregate to compute service coupling. > > That said, there's still lots of missing pieces before placement gets > affinity/anti-affinity support... > Thanks Jay, we are also considering how to leverage the placement aggregates, and if possible, we would like to contribute in this part to make placement work well for mogan :) > > Best, > -jay > > > __________________________________________________________________________ > OpenStack Development Mailing List (not for usage questions) > Unsubscribe: [email protected]?subject:unsubscribe > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev > -- Best Regards, Zhenguo Niu
__________________________________________________________________________ OpenStack Development Mailing List (not for usage questions) Unsubscribe: [email protected]?subject:unsubscribe http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
