Open Stack

Excerpts from Zane Bitter's message of 2015-10-09 17:09:46 +0000:
> On 08/10/15 21:32, Ian Wells wrote:
> >
> >     > 2. if many hosts suit the 5 VMs then this is *very* unlucky,because we should be choosing a host at random from the set of
> >     suitable hosts and that's a huge coincidence - so this is a tiny
> >     corner case that we shouldn't be designing around
> >
> >     Here is where we differ in our understanding. With the current
> >     system of filters and weighers, 5 schedulers getting requests for
> >     identical VMs and having identical information are *expected* to
> >     select the same host. It is not a tiny corner case; it is the most
> >     likely result for the current system design. By catching this
> >     situation early (in the scheduling process) we can avoid multiple
> >     RPC round-trips to handle the fail/retry mechanism.
> >
> >
> > And so maybe this would be a different fix - choose, at random, one of
> > the hosts above a weighting threshold, not choose the top host every
> > time? Technically, any host passing the filter is adequate to the task
> > from the perspective of an API user (and they can't prove if they got
> > the highest weighting or not), so if we assume weighting an operator
> > preference, and just weaken it slightly, we'd have a few more options.
> 
> The optimal way to do this would be a weighted random selection, where 
> the probability of any given host being selected is proportional to its 
> weighting. (Obviously this is limited by the accuracy of the weighting 
> function in expressing your actual preferences - and it's at least 
> conceivable that this could vary with the number of schedulers running.)
> 
> In fact, the choice of the name 'weighting' would normally imply that 
> it's done this way; hearing that the 'weighting' is actually used as a 
> 'score' with the highest one always winning is quite surprising.
> 
> cheers,
> Zane.
> 

There is a more generalized version of this algorithm for concurrent
scheduling I've seen a few times - Pick N options at random, apply
heuristic over that N to pick the best, attempt to schedule at your
choice, retry on failure. As long as you have a fast heuristic and your
N is sufficiently smaller than the total number of options then the
retries are rare-ish and cheap. It also can scale out extremely well.

Obviously you lose some of the ability to micro-manage where things are
placed with a scheduling setup like that, but if scaling up is the
concern I really hope that isnt a problem...

Cheers,
Greg