[mpich-discuss] Can ch3:nemesis:mx use both MX and TCP networks ?

[Audet, Martin]

Hi Scott,

For now this feature is a "nice to have" we don't really need it. It would give us some flexibility for one specific (but important) project.

For the same specific project we would be much more interested in being able to connect with an intercommunicator two independent MPI jobs using MPI_Comm_connect()/MPI_Comm_accept().

Since mpich2-1.1.1p1 a phase was added about this in the RELEASE_NOTES saying that:

 "ch3:nemesis:mx does not support dynamic processes at this time." 

We hope it will be fixed some days. In the mean time, for again the same specific project, we mostly use OpenMPI which allows dynamic process and is very flexible in choosing the fastest communication method between two processes (for the fun I tried one job spread over three clusters where GM, MX, TCP/IP and shared memory were used).

Martin

-----Original Message-----
From: mpich-discuss-bounces at mcs.anl.gov [mailto:mpich-discuss-bounces at mcs.anl.gov] On Behalf Of Darius Buntinas
Sent: October 14, 2009 11:51
To: mpich-discuss at mcs.anl.gov
Subject: Re: [mpich-discuss] Can ch3:nemesis:mx use both MX and TCP networks ?

Martin,

How critical is this feature for you?

-d

On 10/13/2009 11:00 AM, Audet, Martin wrote:
> MPI_Users/MPI_Developers,
> 
>  
> 
> I always thought (but never tested) that the ch3:nemesis:mx channel has
> the ability to independently select either MX or TCP to communicate
> between any pair of MPI processes when shared memory communication isn't
> possible (e.g. the two process run on different nodes).
> 
>  
> 
> However after carefully reading the README file, I fear mpich2 doesn't
> have this flexibility that is, network communication with the
> ch3:nemesis:mx channel is always done using MX (of course network
> communication occurs when shared memory communication isn't possible).
> 
>  
> 
> Can someone clarify this ?
> 
>  
> 
> When shared memory communication isn't possible, does the ch3:nemesis:mx
> channel has the ability to independently select either MX or TCP to
> communicate between two MPI processes ?
> 
>  
> 
> Thanks,
> 
>  
> 
> Martin
>