mpdboot for a cluster

[homng]

Dear all,

I have following lines in my job script:

# Number of cores:

#SBATCH --nodes=8 --ntasks-per-node=8

## Set up job environment

source/site/bin/jobsetup

#startmpd:

mpdboot -n 64

## Run the program:

../bin/sem3dcaf ../input/test_nproc64_sf2.psem

I amtryingto run a coarray program with 64 cores (8cores per node). But I could not correctly start mpd. I am new to intel cluster toolkit. I would be grateful for the suggestion.

Thanks.

[Dmitry_K_Intel2]

Hi homng,

Seems that it would be nice to read Getting Started from Intel MPI Library documentation.

mpdboot create an mpd ring and '-n' says how many nodes will be used. In case of 'mpdboot -n 12' mpd ring will be created using 12 nodes - doesn't matter how many cores on each.

So, you need to change your script:

#startmpd:

mpdboot -n 8

To successfully create an mpd ring you need to have passwodless ssh connection. You can check it by:
$ ssh node_one
From node_one:
$ ssh node_two
and so on.

Usually mpdboot gets list of nodes from mpd.hosts file located in the currect directory, but you can use '-f' option.

mpdtrace shows you a list of nodes in the ring.

I assumed that environment was set up properly.

Regards!
Dmitry

[homng]

Thank you for the suggestion. I went through the documentation but could not really figure out the way to start the mpd from job script. For example, I tried following job script:

#SBATCH --nodes=2 --ntasks-per-node=8

## Set up job environment

source/site/bin/jobsetup

#startmpd:

mpdboot -n 2

But I get the following errors:

totalnum=2 numhosts=1

there are not enough hosts on which to start all processe

totalnum=2 numhosts=1there are not enough hosts on which to start all processes

Possible cause would be that other node is not visible, butI confirmed that I can ssh to all the nodes without password. I can run ordinary MPI program without problem using"mpirun -n 16 test"in the job script. But of course I cannot define the mpd.hosts file because in the big cluster I don't know in advance which nodes will be allocated for my job. I think I am missing some key points!

Your help is greatly appreciated.

Thanks

[homng]

Finally, I think I solved the problem. Thanks.

[Dmitry_K_Intel2]

If you run your application under any job scheduler you need to use 'mpirun -n #_of_processes program' because mpirun understands scheduler's settings readding appropriate environment variables.

In case of co-array, it seems to me that compiler creates a config file which is used by a program itself and you don't need to create an mpd ring. I believe that people from Fortran forum may provide more information about co-array programs invocations.

Regards!
Dmitry