[GE users] mpiexec Hydra integration SGE

reuti reuti at staff.uni-marburg.de
Tue Dec 7 16:19:48 GMT 2010


Am 07.12.2010 um 17:00 schrieb nixter:

> I try to run Intel MPI (1.4.01) based on mpich2. I read that's possible to integrate MPICH2 easly with SGE.
> I used my previous pe like this :
> pe_name            impi_robin
> slots              999
> user_lists         NONE
> xuser_lists        NONE
> start_proc_args    /Softs/Sge/mpi/startmpi.sh -catch_rsh $pe_hostfile
> stop_proc_args     /Softs/Sge/mpi/stopmpi.sh
> allocation_rule    $round_robin
> control_slaves     TRUE
> job_is_first_task  FALSE
> urgency_slots      min
> accounting_summary FALSE
> I execute mpi in my script file like ;
> ..
> mpiexec.hydra -v  -bootstrap rsh  -n 2 ./stress
> the machines file provided by SGE  containes :
> node1
> node2
> I noticed that the application executed only on node1 (2 program)
> It seem like hydra don't care about SGE allocation rules!
> What can i do to have a fully integration with SGE. i.e one process by node ?

you will have to use the mpd startup method for now. It's known not to work with SGE out-of-the-box, as Intel MPI is based on a former version of MPICH2 and their hydra startup is/was experimental at that time. MPICH2 version 1.3.1 will work and out-of-the-box with SGE. A later version of Intel MPI most likely too. The setup of mpd startup is documented here:



A quick check whether the tight integration is built-in, it should list some SGE stuff:

$ strings mpiexec.hydra | grep SGE

-- Reuti

> Thanks in advance.
> nixter.
> ------------------------------------------------------
> http://gridengine.sunsource.net/ds/viewMessage.do?dsForumId=38&dsMessageId=302816
> To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].


To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].

More information about the gridengine-users mailing list