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

Re: [Condor-users] ClassAd for total number of Condor jobs



On Sun, Apr 23, 2006 at 10:46:47PM -0400, Hrant P. Hratchian wrote:
> I'm running Condor across a cluster of 2x SMP machines.  I recently 
> found the TotalLoadAvg ClassAd, which is quite handy.  I'm looking 
> for a class ad that will give the total number of Condor jobs running 
> on a machine.  Does anyone know if one exists or what needs to be 
> done to define one?

There isn't one defined. 

> Basically, I want to use rank so that a new job 
> preferentially starts on a node with 0 jobs running, rather than one 
> with 1 job running, which is the case at present.
> 

You can accomplish that with something like this:

rank =  (vm1_State =!= "CLAIMED") + (vm2_State =!= "CLAIMED")

You will probably have to put STATE into STARTD_VM_EXPRS, I don't
think it's there by default.

However, vm state isn't updated during the negotiation cycle, so 
if a machine has two CPUs and they're both busy, Condor will match them
both (because they both look like their completely unclaimed). You'd
have the same problem if there was an attribute for the number of
currently running jobs on the machine, too.

-Erik