[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Condor-users] about multiple scheduler issues



On Sat, Feb 19, 2005 at 04:40:24PM +0800, Carson Hung wrote:
> Hi,
> 
> Since it is possible to have more than one schedulers/submittors in the
> clusters.
> So I would like to know if it is possible that 2 or more schedulers match
> the same resource to the different jobs if they come from different
> schedulers/submittors.
> 

I think you really mean "if it is possible that 2 or more negotiators match
the same resources to different jobs if they come from different schedds" -
you can't use the term 'scheduler' twice :)

In Condor, it's not really possible for the negotiator to match a resource
twice, because:

A. There's only one negotiator, so it will only match a resource once
per neogotiaition cycle.

B. It is possible for a resource to be matched twice - if it gets matched in
one negotiation cycle, but doesn't update the collector before the next 
negotiation cycle starts. We prevent this from happening in two ways: 
   1. The negotiation CANNOT start until at least 20 seconds have passed since
   the last one. That usually is good enough.
   2. The startd can detect if it's been matched twice, and can recover, if
   the 20 seconds wasn't good enough.

In Condor-G, because there is no claiming (ie the resource doesn't 
automatically tell the collector that it is no longer able to be matched),
matching twice is more of a problem. We ususally build-in some sort of 
claiming when we've deployed matchmaking in the past (see the USCMS writeup
I pointed you at eariler).

If you advertise your resource to more than one collector/negotiator, then
matching twice is a real problem. In regular condor, it's not possible for
a startd to report to two different collectors, so it's not a problem.

-Erik