We had a single unconference session this time at the Ocata summit. The
full etherpad is here:
https://etherpad.openstack.org/p/ocata-nova-summit-unconference
We spent most of the session talking about three newly proposed virt
drivers.
IBM PowerVM
-----------
The PowerVM driver has been getting developed out of the nova tree but
in the openstack namespace for a couple of years now. The team working
on it has been attending summits and meetups providing status and
working on CI. There are some users. There is a spec proposed:
https://review.openstack.org/#/c/381772/
to start adding in some of the minimum required functionality for an
in-tree driver and start running the PowerVM CI on Nova changes.
There was quite a bit of discussion in the room about the driver, like
whether or not it requires cinder/neutron plugins (it doesn't), but
ultimately there was general agreement on the plan to ease the driver in
and iterate on adding feature parity from the out of tree driver to nova
in later releases.
IBM zDPM
--------
This was more socializing the driver and Q&A about it. There is no code
so nothing really to talk about getting in tree. There was a note that
it takes 10 minutes to create a server, which raised some questions.
The direction here was basically follow the PowerVM driver model, which
is work on the driver in the open, work on CI, work on customers, and
attend developer events / stay involved in the Nova community to provide
updates on status.
Huawei FusionCompute
--------------------
This is a driver that Huawei has had for their FusionSphere product,
since around the Folsom release. The code has been made available on
github. It requires a cinder driver but not a neutron ML2 plugin.
The general feeling on this driver was it's not a good fit for Nova
given it's an abstraction to an entire other virtualization management
system, sort of like vCenter or PowerVC. That kind of architecture
raises issues with resource tracking and state management and while yes
the vCenter driver is in nova, the architecture it relies on has been
problematic and we don't really want to make the same mistake with more
virt drivers.
There was some feedback on the other side of this to the effect that if
there are customers / users of the FusionCompute driver, why would we
not allow it into Nova - and if there are limitations with it then it's
up to the vendor to deal with them, or the users to have to live with
them. This is a debate that isn't likely to go away anytime soon given
the platform9/omni demo at the summit which is essentially a virt driver
for AWS.
So while it doesn't look probable for this driver to get into Nova, at
least anytime soon, it is in Huawei's best interest to develop the
driver in the open and work on CI.
Generalize PCI device as Host device
------------------------------------
This was a spec mentioned during the session:
https://review.openstack.org/#/c/385810/
The intent is to generalize host PCI devices are modeled in Nova so that
new types can be built into that model for new support, like for vGPUs.
There wasn't a lot of discussion on this during the session, and not all
of the best people to weigh in on this were in the room, so the
direction was to move discussion to the spec and get the
SR-IOV/PCI/resource provider people like lbeliveau, jaypipes and moshele
to review the spec.
--
Thanks,
Matt Riedemann
__________________________________________________________________________
OpenStack Development Mailing List (not for usage questions)
Unsubscribe: [email protected]?subject:unsubscribe
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
