Adam Lawson wrote:
> I have a quick question: how is Amazon doing this? When choosing a next
> path forward that reliably scales, would be interesting to know how this
> is already being done.
Well, those who know probably would be sued if they told.
Since they have a limited set of instance types and very limited
placement options, my bet would be that they do flavor-based scheduling
("let compute nodes grab node reservation requests directly
out of flavor based queues based on their own current observation of
their ability to service it" in Clint's own words).
This is the most efficient way to scale: you no longer rely on a
specific scheduler trying to keep an up-to-date view of your compute
nodes resource availability. As long as you are ready to abandon fancy
placement features, you can get simple, reliable and scalable
(non-)scheduling.
Personally as we explore the options we have in that space, I'd like to
consider options that still enable us to plug such a no-scheduler
solution without too much trouble. Just for those of us who are ready to
make that trade-off :)
--
Thierry Carrez (ttx)
__________________________________________________________________________
OpenStack Development Mailing List (not for usage questions)
Unsubscribe: openstack-dev-requ...@lists.openstack.org?subject:unsubscribe
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
