On Tue, Aug 2, 2016 at 7:56 AM, Edric Ellis <Edric.Ellis@mathworks.co.uk> wrote:

Thanks for that Jeff. Are you referring to this discussion? http://lists.mpich.org/pipermail/devel/2016-July/000717.html


Yes
 

If so, that shares some similarity with what I’m trying to recreate in MPICH. However, our code using MPICH2 makes no use whatsoever of mpiexec when in “connect/accept” mode, and I changing this would be rather invasive.

 


I agree that this should work, but as the ticket I shared indicates, it is not implemented right now.
 

It seems as though that open issue has been open for quite a while now, with no scheduled release for when it might be addressed – so am I correct in understanding that there are no known workarounds, and that essentially my only option for now is to stick with MPICH2?

 


You could also try Open-MPI 2.0 (https://www.open-mpi.org/community/lists/announce/2016/07/0085.php).  They fixed a bug related to this reported by the same person who reported the MPICH bug, so perhaps it works for your usage.

Normally, I do not recommend Open-MPI, because it has been full of bugs related to everything I care about for many years, but at least in this one specific case, it is possible they have the upper hand.

Jeff
 

Cheers,

Edric.

 

From: Jeff Hammond [mailto:jeff.science@gmail.com]
Sent: Monday, August 01, 2016 6:43 PM
To: MPICH <discuss@mpich.org>
Subject: Re: [mpich-discuss] Singleton init configuration for MPICH

 

It's an open issue: https://trac.mpich.org/projects/mpich/ticket/1074.

 

There was some discussion on the devel list as of last week that might be of interest to you.

 

Jeff

 

On Mon, Aug 1, 2016 at 8:47 AM, Edric Ellis <Edric.Ellis@mathworks.co.uk> wrote:

Hi,

 

I need to come up with an MPICH build that allows singleton-init processes to use MPI_Comm_connect/accept to get connected. Our MPICH2 build uses the “smpd” process manager to achieve this. However, now that we’re upgrading to MPICH (3.1.4), I notice that the “smpd” process manager no longer exists. Is there a combination of process manager and PMI that allows connect/accept for singleton-init processes? I’ve tried a few things and cannot seem to find a combination that works. Here are the things I’ve tried:

 

1.       Hydra with pmi=simple – fails when attempting to launch hydra’s mpiexec (I don’t have the message to hand, but my recollection is that hydra’s mpiexec didn’t accept “-pmi_args”)

2.       Hydra with pmi=pmi2/simple – cannot be built

3.       Gforker with pmi=pmi2/simple – required a couple of tweaks before it could be built, and then fails in PMPI_Init_thread complaining “Message truncated”

 

Is there some other combination that might work? Ideally, the singleton-init processes wouldn’t need to use mpiexec at all (as was the case with our MPICH2 build using smpd).

 

Cheers,

Edric.

 


_______________________________________________
discuss mailing list     discuss@mpich.org
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/discuss



 

--


_______________________________________________
discuss mailing list     discuss@mpich.org
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/discuss



--
Jeff Hammond
jeff.science@gmail.com
http://jeffhammond.github.io/