On 24/11/15 19:14, John Trowbridge wrote:
Howdy folks.
How feasible do we think $subject is?
It can be done, but its not going to be simple. The challenge is mostly around what the other projects are doing.
For RDO-Manager jobs it is pretty much near zero chance of success on a machine with 16G of RAM. This is especially true for our HA jobs. I made a bit of a hack to our centosci provisioner in khaleesi[1] to throw out all hufty chassis nodes since they only have 16G RAM. This is slightly wasteful though, and would be better to be something that could be selected via the API.
What we can do for now, is take out hufty.ci from the mix completely and replace it with another pool that has the 32gb ram spec - and keep hufty.ci for failover, testing, dev and reserved nodes setup. so from the api side you will not see nodes from hufty.ci allocated anymore.
Thanks, John Trowbridge
[1] https://github.com/redhat-openstack/khaleesi/commit/14d19c317c0a3fb6b551f4b1...
please dont do this! note that there is only a small set ( typically 10 to 12 ) machines in the warm-cache at any given time for c7/x86_64 and they can all be from the same pool ( hufty even! )
If we need to implement something of this nature, I would prefer to do this on the API service side of things, maybe map it from the apikey:<desired config>; that way it would be able to ensure ( or try to! ) that there are some machines always in a cache'd stage, in a pool thats usable for that apikey.
For now, we are able to take hufty.ci out of the loop, so no new allocations are made from that pool.
regards