Open MPI logo

Open MPI User's Mailing List Archives

  |   Home   |   Support   |   FAQ   |  

This web mail archive is frozen.

This page is part of a frozen web archive of this mailing list.

You can still navigate around this archive, but know that no new mails have been added to it since July of 2016.

Click here to be taken to the new web archives of this list; it includes all the mails that are in this frozen archive plus all new mails that have been sent to the list since it was migrated to the new archives.

Subject: Re: [OMPI users] Optimal distribution of low and high CPU processes on a heterogeneous cluster
From: Jeff Squyres (jsquyres_at_[hidden])
Date: 2008-05-05 09:28:25


On Apr 25, 2008, at 2:27 PM, Roopesh Ojha wrote:

> For our application, some processes (manager and the datastreams)
> won't
> take much CPU power while others will. Our application first assigns
> the
> non-compute bound processes and then the high CPU power ones.
> Clearly we
> would like to be able to assign the former processes to our two slower
> machines. How can we get MPI to achieve this?

Sorry for the delay in replying.

I'm not entirely clear on exactly what you want to do. I think you're
saying that you have 4 nodes:

fastnode1
fastnode2
slownode1
slownode2

And you want to schedule some slow procs and fast procs. I think you
should be able to do something like:

mpirun -np 2 --host slownode1,slownode2 slowproc : \
        -np 2 --host fastnode1,fastnode2 fastproc

> What is the best way to
> use
> slot and max-slot commands in the hostfile to optimize performance?

These options are not really about performance; they're used for
distributing processes across nodes. In your case, those are tightly
integrated, but note that they're really different things.

*Never* tell OMPI that there are more slots than are actually
available on a host.

> How
> does max-slot behave when some node is listed more than once in the
> hostfile?

IIRC, it's additive.

> Is there any way to control where processes are deployed apart from
> the
> basic bynode or byslot round robin schedule?

In the v1.2 series, no. In the v1.3 series, some new patterns were
just implemented within the last week or three -- docs have not been
written up yet, but Ralph Castain described some of them on the devel
list. Have a look through those archives on the web for more detail
(I'd get the details wrong if I tried to explain them).

-- 
Jeff Squyres
Cisco Systems