[mpich-discuss] Hydra: prefer localhost instead of first in host file

[Torquil Macdonald Sørensen]

Hi Pavan!

Ok, I've implemented an option in my simulation script to specify use of the 
hosts file or not, so I'm choosing it on a case by case basis. It seems to work 
fine.

Thanks!
Torquil Sørensen

On 27/03/11 16:45, Pavan Balaji wrote:
> Hi Torquil,
>
> mpiexec -n 4 prog.exe (without any host file option) will always run
> processes on the local node.
>
> I'm assuming you are setting the host file asking Hydra to use hostA,
> followed by hostB, followed by hostC. Hydra is doing exactly that. If
> you don't want Hydra to do that, you should remove the hostfile option.
>
> -- Pavan
>
> On 03/27/2011 09:07 AM, Torquil Macdonald Sørensen wrote:
>> Hi!
>>
>> I'm in an environment with multiple computers sharing the same
>> filesystem.
>>
>> I'm using the latest Mpich2 with hydra and a host file that looks like
>>
>> hostA:4
>> hostB:4
>> hostC:4
>>
>> The host file is the same on all the computers, because it is on a shared
>> filesystem.
>>
>> Sometimes I want to run with 12 processes and in that case this works
>> fine.
>>
>> However, sometime I want only 4 processes, and I would like them to
>> run on the
>> computer I'm currently logged into (not always hostA). E.g. I want
>> "mpiexec -n 4
>> prog.exe" to start 4 processes on hostB if I run this command from hostB.
>>
>> That way, I can e.g. run three different simulations, each of 4
>> processes, each
>> on its own computer, just by executing "mpiexec -n 4 prog.exe" on each
>> individual computer.
>>
>> With earlier versions of Mpich2 this seemed to work as I wished (not
>> using Hydra).
>>
>> Is there a way of making Mpich2 prefer to start the first 4 process on
>> the local
>> machine instead of starting them on the computer listed first in the
>> host file?
>>
>> I'v been reading the Hydra documentation, obviously without finding
>> out if this
>> can be accomplished.
>>
>> Best regards
>> Torquil Sørensen
>> _______________________________________________
>> mpich-discuss mailing list
>> mpich-discuss at mcs.anl.gov
>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>