[openstack-dev] [all] [nova] [placement] placement below or beside compute after extraction?

Stephen Finucane sfinucan at redhat.com
Mon Aug 27 11:26:00 UTC 2018


On Sat, 2018-08-25 at 08:08 +0800, Alex Xu wrote:
> 
> 
> 2018-08-18 20:25 GMT+08:00 Chris Dent <cdent+os at anticdent.org>:
> > On Fri, 17 Aug 2018, Doug Hellmann wrote:
> > 
> > > If we ignore the political concerns in the short term, are there
> > > other projects actually interested in using placement? With what
> > > technical caveats? Perhaps with modifications of some sort to
> > > support
> > > the needs of those projects?
> > > 
> >  
> > I think ignoring the political concerns (in any term) is not
> > possible. We are a group of interacting humans, politics are always
> > present. Cordial but active debate to determine the best course of
> > action is warranted.
> > 
> > (tl;dr: Let's have existing and potential placement contributors
> > decide its destiny.)
> > 
> > Five topics I think are relevant here, in order of politics, least
> > to most:
> > 
> > 1. Placement has been designed from the outset to have a hard
> > contract between it and the services that use it. Being embedded
> > and/or deeply associated with one other single service means that
> > that contract evolves in a way that is strongly coupled. We made
> > placement have an HTTP API, not use RPC, and not produce or consume
> > notifications because it is supposed to be bounded and independent.
> > Sharing code and human management doesn't enable that. As you'll
> > read below, placement's progress has been overly constrained by
> > compute.
> > 
> > 2. There are other projects actively using placement, not merely
> > interested. If you search codesearch.o.o for terms like "resource
> > provider" you can find them. But to rattle off those that I'm aware
> > of (which I'm certain is an incomplete list):
> > 
> > * Cyborg is actively working on using placement to track FPGA
> >   e.g., https://review.openstack.org/#/c/577438/
> > 
> > * Blazar is working on using them for reservations:
> >   
> > https://review.openstack.org/#/q/status:open+project:openstack/blazar+branch:master+topic:bp/placement-api
> > 
> > * Neutron has been reporting to placement for some time and has
> > work
> >   in progress on minimum bandwidth handling with the help of
> >   placement:
> >   
> > https://review.openstack.org/#/q/status:open+project:openstack/neutron-lib+branch:master+topic:minimum-bandwidth-allocation-placement-api
> > 
> > * Ironic uses resource classes to describe types of nodes
> > 
> > * Mogan (which may or may not be dead, not clear) was intending to
> >   track nodes with placement:
> >   
> > http://git.openstack.org/cgit/openstack/mogan-specs/tree/specs/pike/approved/track-resources-using-placement.rst
> > 
> > * Zun is working to use placement for "unified resource
> > management":
> >   
> > https://blueprints.launchpad.net/zun/+spec/use-placement-resource-management
> > 
> > * Cinder has had discussion about using placement to overcome race
> >   conditions in its existing scheduling subsystem (a purpose to
> >   which placement was explicitly designed).
> > 
> > 3. Placement's direction and progress is heavily curtailed by the
> > choices and priorities that compute wants or needs to make. That
> > means that for the past year or more much of the effort in
> > placement
> > has been devoted to eventually satisfying NFV use cases driven by
> > "enhanced platform awareness" to the detriment of the simple use
> > case of "get me some resource providers". Compute is under a lot of
> > pressure in this area, and is under-resourced, so placement's
> > progress is delayed by being in the (necessarily) narrow engine of
> > compute. Similarly, computes's overall progress is delayed because
> > a
> > lot of attention is devoted to placement.
> > 
> > I think the relevance of that latter point has been under-estimated
> > by the voices that are hoping to keep placement near to nova. The
> > concern there has been that we need to continue iterating in
> > concert
> > and quickly. I disagree with that from two angles. One is that we
> > _will_ continue to work in concert. We are OpenStack, and
> > presumably
> > all the same people working on placement now will continue to do
> > so,
> > and many of those are active contributors to nova. We will work
> > together.
> > 
> > The other angle is that, actually, placement is several months
> > ahead
> > of nova in terms of features and it would be to everyone's
> > advantage if
> > placement, from a feature standpoint, took a time out (to extract)
> > while nova had a chance to catch up with fully implementing shared
> > providers, nested resource providers, consumer generations,
> > resource
> > request groups, using the reshaper properly from the virt drivers,
> > having a fast forward upgrade script talking to PlacementDirect,
> > and
> > other things that I'm not remembering right now. The placement side
> > for those things is in place. The work that it needs now is a
> > _diversity_ of callers (not just nova) so that the features can
> > been
> > fully exercised and bugs and performance problems found.
> > 
> > The projects above, which might like to--and at various times have
> > expressed desire to do so--work on features within placement that
> > would benefit their projects, are forced to compete with existing
> > priorities to get blueprint attention. Though runways seemed to
> > help
> > a bit on that front this just-ending cycle, it's simply too dense a
> > competitive environment for good, clean progress.
> > 
> > 4. While extracting the placement code into another repo within the
> > compute umbrella might help a small amount with some of the
> > competition described in item 3, it would be insufficient. The same
> > forces would apply.
> > 
> > Similarly, _if_ there are factors which are preventing some people
> > from being willing to participate with a compute-associated
> > project,
> > a repo within compute is an insufficient break.
> > 
> > Also, if we are going to go to the trouble of doing any kind of
> > disrupting transition of the placement code, we may as well take as
> > a big a step as possible in this one instance as these
> > opportunities
> > are rare and our capacity for change is slow. I started working on
> > placement in early 2016, at that time we had plans to extract it to
> > "it's own thing". We've passed the half-way point in 2018.
> > 
> > 5. In OpenStack we have a tradition of the contributors having a
> > strong degree of self-determination. If that tradition is to be
> > upheld, then it would make sense that the people who designed and
> > wrote the code that is being extracted would get to choose what
> > happens with it. As much as Mel's and Dan's (only picking on them
> > here because they are the dissenting voices that have showed up so
> > far) input has been extremely important and helpful in the
> > evolution
> > of placement, they are not those people.
> > 
> > So my hope is that (in no particular order) Jay Pipes, Eric Fried,
> > Takashi Natsume, Tetsuro Nakamura, Matt Riedemann, Andrey Volkov,
> > Alex Xu, Balazs Gibizer, Ed Leafe, and any other contributor to
> > placement whom I'm forgetting [1] would express their preference on
> > what they'd like to see happen.
> 
> Sorry, I didn't read all the reply, compare to 70 replies, I prefer
> to review some specs...English is heavy for me.
> 
> I'm not very care about the extraction. But in the currently
> situation, I think placement contributors and nova contributors still
> need work to together, the resharp API is an example. So whatever we
> extract the placement or not, pretty sure nova and placement should
> work together.
> 
> And really hope we won't have separate room in the PTG for placement
> and nova..I don't want to make a hard choice to listen which one...I
> already used to stay at one spot in a week now.

What he said, minus the "English is heavy" bit. The only thing I care
about is making sure the odd nova-placement'y thing I might care about
(vGPU and generic "devices" at large, maybe a future version of NUMA-
aware vSwitches) don't get significantly more difficult to implement
post-whatever it is we end up doing. Once that constraint is satisfied,
it's all good.

Now, best get started on those spec reviews, I guess...

Stephen

> > At the same time, if people from neutron, cinder, blazar, zun,
> > mogan, ironic, and cyborg could express their preferences, we can
> > get
> > through this by acclaim and get on with getting things done.
> > 
> > Thank you.
> > 
> > [1] My apologies if I have left you out. It's Saturday, I'm tried
> > from trying to make this happen for so long, and I'm using various
> > forms of git blame and git log to extract names from the git
> > history
> > and there's some degree of magic and guessing going on.
> > 
> > 
> > ___________________________________________________________________
> > _______
> > OpenStack Development Mailing List (not for usage questions)
> > Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsu
> > bscribe
> > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev





More information about the OpenStack-dev mailing list