Hi,

Am 17.12.2014 um 10:06 schrieb Hairul Ikmal M. Fuzi:

>> export MPI_REMSH=rsh
>> 
> If I am using 'ssh', should I set as MPI_REMSH=ssh ?

You mean the wrapper you use is called `ssh` - or where are you using `ssh`?

The example wrapper in $SGE_ROOT/mpi is called `rsh`, therefore this name needs 
to be set for Platform MPI. To emphasize this chain of commands:

You can define "MPI_REMSH=foobar". The mpi wrapper which must be accessible 
during the execution of the parallel job in some directory like $SGE_ROOT/mpi 
needs to be called `foobar` then too (it's an arbitrary name only) [this link 
in $TMPDIR to point to $SGE_ROOT/mpi is created by startmpi.sh which must be 
attached to the PE's start_proc_args]. Next the `foobar` will call `qrsh 
-inherit ... `which will use the set up communication method in SGE's 
configuration `qconf -sconf`, being it builtin`, `ssh` or `rsh`.


>> the usual wrapper for MPI applies, I.e. a Platform MPI  calls `rsh`, this is 
>> caught and routed to `qrsh -inherit ...`. It's also necessary to use a 
>> reformatted machine file for `mpiexec` which is generated by 
>> $SGE_ROOT/mpi/startmpi.sh called "$TMPDIR/machines" (Platform MPI can read 
>> the MPICH(1) file format of the hostlist).
>> 
> 
> I saw some example job submissions utilizes the switch '-pe XXXX' when
> running MPI job scripts with 'qsub'. Should I start with creating a
> relevant 'parallel environment' configuration first for Platform MPI?

Yes. The PE is mandatory in any case.

-- Reuti


>> The Platform MPI is also available in a free community edition, in case you 
>> want to set it up beforehand:
>> 
>> http://www.ibm.com/developerworks/downloads/im/mpi/
>> 
> 
> This is a very useful information, thank you!
> 
> 
> -Ikmal
> 
> 
> On Wed, Dec 17, 2014 at 4:36 PM, Reuti <[email protected]> wrote:
>> 
>> Hi,
>> 
>> Am 17.12.2014 um 09:01 schrieb Hairul Ikmal M. Fuzi:
>> 
>>> I am one of the user of GridEngine (GE). Previously we use GE only for 
>>> serial jobs that runs independently.
>>> 
>>> Soon, we will have 'parallel' application that will use MPI (Platform MPI, 
>>> I think, based on product description). I tried to do a quick read on how 
>>> we could use GE and Platform MPI together.
>>> 
>>> I found out the documentation to do tight integration (for various MPI 
>>> libraries) is available at 
>>> http://gridscheduler.sourceforge.net/howto/GridEngineHowto.html
>>> 
>>> However, it does not include Platform MPI.
>>> 
>>> Could we do a tight integration with GE and Platform MPI ?
>> 
>> Yes. After setting in your jobscript:
>> 
>> export MPI_REMSH=rsh
>> 
>> the usual wrapper for MPI applies, I.e. a Platform MPI  calls `rsh`, this is 
>> caught and routed to `qrsh -inherit ...`. It's also necessary to use a 
>> reformatted machine file for `mpiexec` which is generated by 
>> $SGE_ROOT/mpi/startmpi.sh called "$TMPDIR/machines" (Platform MPI can read 
>> the MPICH(1) file format of the hostlist).
>> 
>> The Platform MPI is also available in a free community edition, in case you 
>> want to set it up beforehand:
>> 
>> http://www.ibm.com/developerworks/downloads/im/mpi/
>> 
>> -- Reuti
>> 
>> PS: Some tools we use like ADF (reads $PE_HOSTFILE) and Turbomole (reads 
>> $TMPIDIR/machines) take care of the hostfile already on their own.
>> 
>> 
>>> If yes, which document should I follow to do the integration? Or is there 
>>> any other reference that I could use?
>>> 
>>> Thank you in advance.
>>> 
>>> 
>>> -Ikmal
>>> 
>>> _______________________________________________
>>> users mailing list
>>> [email protected]
>>> https://gridengine.org/mailman/listinfo/users
>> 
> 


_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users

Reply via email to