[mpich-discuss] SGE & Hydra Problem

[Pavan Balaji]

On 09/14/2010 02:42 PM, Reuti wrote:
>> On 09/14/2010 10:10 AM, Reuti wrote:
>>> did you also recompile all applications with the latest version of
>>> MPICH2, so that binaries and mpiexec are from the same version?
>>
>> Actually, you don't need to recompile the application at all. If the
>> application was compiled with any version of any MPICH2 derivative,
>> it's sufficient.
>
> thanks for clarification - I wasn't aware of it. In the past I
> struggled a little bit with the different compilation methods of
> MPICH2 (mpd/smpd/smpd with rsh/gforker) and my conclusion was, they
> must always match the compilation of all involved parts.

Well, let me clarify this. There are a few different implementations of 
the PMI wire protocol -- simple PMI, smpd PMI and slurm PMI are the 
three common ones. If you build MPICH2 or derivative with one of them, 
you can use all process managers that support the same PMI wire protocol.

simple PMI is the default when you build MPICH2 or its derivatives. So 
you can use any process manager that supports simple PMI's wire protocol 
(including Hydra, mpd, gforker, remshell, or the process manager that's 
shipped with MVAPICH: mpirun_rsh, or the process manager distributed 
from OSC: OSC mpiexec).

smpd PMI is only supported by the SMPD process manager, and slurm PMI is 
only supported by the SLURM process manager.

So, as long as you build MPICH2 or derivatives with simple PMI, you 
should be able to use Hydra.

> BTW: will the other methods be dropped over time, when Hydra is the to
> be used one in the future?

Correct.

  -- Pavan

-- 
Pavan Balaji
http://www.mcs.anl.gov/~balaji