Re: [Condor-users] Using Condor as a substitute for OpenPBS

[Jaime Frey <jfrey@xxxxxxxxxxx>]

On Oct 27, 2005, at 5:25 AM, Angel de Vicente wrote:

> Marko Kääramees writes:
>
> > > despite being a fan of Condor (and administering a 200 machines  
> > > pool), I would
> > > not recommend it as a substitute of OpenPBS for a cluster. I  
> > > think you would be
> > > better off trying Torque
> >
> > Could you explain a little please why do you prefer Torque as a  
> > cluster
> > manager if there is anything more than originating from OpenPBS.
>
> just a few things from when I had to take the decision on the  
> scheduler for our
> cluster:
>
> * Condor (at least the stable version we are using) can only work  
> with MPICH
>   1.2.4, and we need 1.2.6, MPICH2, LAM/MPI.

This is something we're very close to fixing.

> * Our nodes are SMP machines, so when running a program on 16 CPUS  
> it is not the
>   same to use 16 nodes (1 cpu in each node), 4 nodes (4 cpus in  
> each node), or
>   other combinations. I looked at ways in which this could be  
> accomplished with
>   Condor, but as matching expressions didn't work with regular  
> expressions, I
>   didn't find a way to get this working.

By default, Condor will advertise each cpu as an individual resource.  
An MPI job will get 16 cpus, not 16 nodes.

> * Condor is great at CPU harvesting, but in a dedicated cluster it  
> is probably a
>   bit too heavyweight: it monitors things like keyboard activity,  
> mouse
>   activity, etc. which are not important in our cluster, as the  
> nodes are
>   dedicated and not accessible to users, unless they submit a job  
> through the
>   queueing system.

Monitoring of keyboad, mouse, and load shouldn't produce any  
noticeable load on the machine.

> * In combination with Maui, Torque offers a very fine control on usage
>   policies. For instance, in our cluster we have settings that  
> guarantee a group
>   of users 40% of CPU time, and makes sure that no job that needs  
> the whole
>   cluster for a period of more than 7 days will be accepted. There  
> are probably
>   ways in which you can accomplish something similar with Condor,  
> but the
>   emphasis of the system is not there, so you would have to go a  
> little bit off
>   your way to get this working.
>
> Actually if these points could be solved with a newer version of  
> Condor, I would
> gladly swap schedulers, so that I could incorporate the cluster to  
> our Condor
> pool, but for the time being I think Torque/Maui is a better option.

+----------------------------------+---------------------------------+
|            Jaime Frey            |  Public Split on Whether        |
|        jfrey@xxxxxxxxxxx         |  Bush Is a Divider              |
|  http://www.cs.wisc.edu/~jfrey/  |         -- CNN Scrolling Banner |
+----------------------------------+---------------------------------+