Re: [Condor-users] Gang-Matching in condor and MPI

[Mário Costa <atritoman@xxxxxxxxx>]

On Sat, 12 Mar 2005 13:06:33 -0600, Erik Paulson <epaulson@xxxxxxxxxxx> wrote:
> On Sat, Mar 12, 2005 at 12:16:21PM +0100, M?rio Costa wrote:
> > Hi, Can any tell-me if gang-matching is going to be added to condor in
> > a near future ?
> >
> 
> Not anytime soon.
> 
> > I needed to specify that a parallel job (MPI), for example that
> > requires 4 processors should use 2 processors in a condor pool, and
> > other 2 in another condor poll that are connected by flocking ...
> >
> 
> As of 6.7.5, you are able to specify different requirements for different
> nodes in an MPI job. They all still must be bound to one dedicated
> scheduler, you cannot split an MPI job across pools.
> 

I'm submiting jobs from one dedicated scheduler, so regarding to that
there is no problem ...

> If you desperately wanted to, you could submit 4 jobs that were 'sshd'
> or something similar that would let you run jobs underneath them. (In
> fact, what I would probably do is submit condor_startds, ala glide-in,
> that are pre-configured to join another dedicated scheduler. That way
> you could build an virtual, overlay pool and then use the dedicated
> scheduler to run MPI jobs. Of course, the "dedicated" in dedicated
> scheduler means that it doesn't expect jobs to be preempted, so if
> your glide-in gets preempted everything falls apart...)
> 

I just wanted to submit the job, and force some of the nodes to flock
in to another pool ...

I've looked at the flocking mechanisms, but I havent undestud how can
one specify policies about the the jobs we are wiling to run in our
condor pool ... Is it just about the hosts we alow jobs to flock from
and to?

Or can we have some more control, like filtring jobs with some caracteristics ?
Is it implemented like described in "A worldwide flock of condors:
Load Sharing among workstation clusters", By D.H. Epema, M. Livny ...
??

With gang matching this would be mutch more simple ... still I have
another question, we can specify diferent sets of restricions for
diferent sets of nodes in a MPI job, is there any way that one could
reference the classad from a previous match, to the next, folowing the
queue statments as if they where diferent scopes, binding the previous
matched nodes with a vareable, say SUPER ??

someting like this

######################################
## MPI example submit description file
## with multiple procs
######################################
universe = MPI
...
machine_count = 1
...
requirements = ( TARGET.pool == "local pool id")
queue

requirements = ( SUPER.TARGET.pool != TARGET.pool)
machine_count = 3
queue

Where super, stands for the classad of the previous mached node(s),
its someting similar to gang matchig in a way, because there are a
king of binding betewn previous matches ... this can not be done right
now ??


Thanks in advace.

> -Erik
> _______________________________________________
> Condor-users mailing list
> Condor-users@xxxxxxxxxxx
> https://lists.cs.wisc.edu/mailman/listinfo/condor-users
>