Open Stack

On Apr 14, 2011, at 9:40 AM, Soren Hansen wrote:

>> When scheduling, the instance has not been created yet. We have to make decisions on where the instance will ultimately reside on a number of factors:
>> 
>> 1. the capabilities of the host hypervisor (the Compute node)
>> 2. the current load the host is under
>> 3. the distribution policies of the customer (ie. "don't put all my instances on the same host/zone, etc")
> 
> I still question the usefulness of nr. 2. A host that is almost
> completely idle right now might be under tremendous pressure a minute
> from now and vice versa. Even if we had useful statistics (and trend
> analysis) of the existing VM's on the host, we have no data at all on
> the requirements of the instance we're about to launch.
> 
> It seems to me that the current load only affects the time it will
> take to spin up the new instance, and as such says nothing about its
> long term (or even medium or short term!) suitability as the new
> instance's home.


	Such a consideration would not be the primary selection criterion, but would be more of a tie breaker. E.g., given a number of hosts that appear to be roughly equally good choices for a new instance, why not weight those that are currently under a heavy load so that they are probably not selected?

	This is the current way that the Slicehost codebase works: the load of a host influences how likely it is to be selected, but it is not the main factor.


-- Ed Leafe