Review wsgi process/thread config

[Oliver Walsh]

Based on http://lists.openstack.org/pipermail/openstack-operators/2018-March/015044.html...

Looking at the tripleo/puppet config we appear to set procs to a function of the number of CPUs, and threads to 1 (everywhere), which may not scale well.

[Alex Schultz (alex-schultz)]

In the past we've been told that threads should be 1 and the process count should be tuned. It depends on the service i guess. For processes we use the os_worker fact that uses a sliding scale with a cap so folks don't end up with 48+ processes for every service running

[Oliver Walsh (owalsh)]

Ok, <sigh> so we may need to review each individual service...

Just looking at nova-placement for now:

(Chris & Sylvain, any suggestions for the defaults are much appreciated)

The current default in the RDO nova placement rpm is procs = 3, threads = 1.
The current default in puppet-nova (via puppet-openstacklib) is procs = min(max(#cpus/2, 2), 12), threads = 1.

The number of procs are perhaps a bit on the conservative side in both cases, but not ridiculous as a default IMO.

I'd be inclined to change threads to the mod_wsgi default (15) on master at least, unless we know of a very good reason not to.

[Chris Dent (cdent)]

I can't really comment on the defaults for anything other than placement, I wish I could but, time, energy, etc.

On placement the reason it's important to have a pretty hefty base is that every nova-compute process will talk to placement a few times per minute. Get the number of nova-computes up in the high numbers and as you might expect things will start to get crowded.

In my experiments, placement works fine when threads are available. That is: I'm not aware of it experiencing the problems that apparently keystone did at some point when it was configured to use both threads and processes.

I agree with your inclination Oliver: especially if you are able to get good feedback from people if things go sideways.

[Marios Andreou (marios-b)]

Bug status has been set to 'Incomplete' and target milestone has been removed due to inactivity. If you disagree please re-set these values and reach out to us on freenode #tripleo

[Takashi Kajinami (kajinamit)]

Because we haven't seen any activity about this bug for long, I'll close this as WON'T FIX.
To move this forward, we first need actual learning/feedback from real deployments.