[openstack-dev] [tripleo][heat] Heat memory usage in the TripleO gate: Ocata edition

Steven Hardy shardy at redhat.com
Wed Mar 15 11:35:28 UTC 2017

On Tue, Mar 14, 2017 at 10:21:54PM -0400, Emilien Macchi wrote:
> On Tue, Mar 14, 2017 at 4:06 PM, Zane Bitter <zbitter at redhat.com> wrote:
> > Following up on the previous thread:
> >
> > http://lists.openstack.org/pipermail/openstack-dev/2017-January/109748.html
> >
> > Here is the latest data, which includes the Ocata release:
> >
> > https://fedorapeople.org/~zaneb/tripleo-memory/20170314/heat_memused.png
> >
> > As you can see, there has been one jump in memory usage. This was due to the
> > TripleO patch https://review.openstack.org/#/c/425717/
> Since Contrail is optional when deploying TripleO, is there an easy
> way to disable the endpoints by default and activate them only when we
> enable Contrail services?

Not any easy way at the moment AFAIK - I think we need a way to fix this
though, I raised a bug:


I think there are a few ways we could potentially rework things to
dynamically generate the endpoint_map.yaml, but the easiest (and least
expensive from a heat memory perspective) is probably to use jinja2

If we try to do the transformation in heat we have a chicken/egg
problem (the EndpointMap is created before the per-role *ServiceChain
resources in overcloud.j2.yaml, so we can't do the normal pattern of
composing lists of per-service things from *ServiceChain), which is
unfortunate, but I can't currently think of a good way around it.

> > Unlike previous increases in memory usage, I was able to warn of this one in
> > the advance, and it was deemed an acceptable trade-off. The reasons for the
> > increase are unknown - the addition of more stuff to the endpoint map seemed
> > like a good bet, but one attempt to mitigate that[1] had no effect and I'm
> > increasingly unconvinced that this could account for the magnitude of the
> > increase.
> >
> > In any event, memory usage remains around the 1GiB level, none of the other
> > complexity increases during Ocata have had any discernible effect, and Heat
> > has had no memory usage regressions.
> >
> > Stay tuned for the next exciting edition, in which I try to figure out how
> > to do more than 3 colors on the plot.

Thanks for the analysis Zane, it's much appreciated!


More information about the OpenStack-dev mailing list