eciencia: Middleware/MpiStart/UserDocumentation


mpi-start is an abstraction layer that offers a unique interface to start parallel jobs with different execution environments implementations. It provides support for different MPI implementations.


Normally users do not need to install mpi-start. However if they want to use it in a site without an existing installation, the recommendation is to create a tarball installation that can be transferred in the input sandbox of the job.

In order to create a tarball installation, get the source code and do the following:

$ make tarball

This will create a mpi-start-X.Y.Z.tar.gz (with X.Y.Z being the version of mpi-start) that contains all that is needed for the execution of jobs. In your job script unpack the tarball and set the I2G_MPI_START environment variable to $PWD/bin/mpi-start.


mpi-start can be controlled via environment variables or command line switches, most configuration dependent paramenters are automatically detected by mpi-start and do not need to be specified by the user. The following command line will be enough to run the application with the site defaults:

$ mpi-start application [application arguments ...]

Command Line Options

show help message and exit
show mpi-start version
-t mpi_type

use mpi_type as MPI implementation

be verbose
include debug information
include full trace
-pre hook

use hook as pre-hook file

-post hook

use hook as post-hook file

-pcmd cmd

use cmd as pre-command

-npnode ''n''

start n processes per node

start 1 process per node
-np ''n''
start exactly 'n' processes
-i file

use file as standard input file

-o file

use file as standard output file

-e file

use file as standard error file


define variable VAR with optional VALUE for the application's environment (will not be seen by mpi-start!)


define variable VAR with VALUE

optional separator for application and arguments, after this, any arguments will be considered the application to run and its arguments

For example, the following command line would start /bin/hostname 3 times for available node using Open MPI:

$ mpi-start -t openmpi -npnode 3 -- /bin/hostname

Environment Variables

Prior to version 1.0.0 mpi-start only used environment variables to control its behavior. This is still possible, although command line arguments will override the environment variables defined. Next table shows the complete list of variables, with the command line options that can be used to set them:


cmd line option



The application binary to execute.


The command line parameters for the application



The name of the MPI implementation to use.



This variable can be set to a script which must define the pre_run_hook function. This function will be called after the MPI support has been established and before the internal pre-run hooks. This hook can be used to prepare input data or compile the program.



This variable can be set to a script which must define the post_run_hook function. This function will be called after the mpirun has finished.



Set to 1 to turn on the additional output.



Set to 1 to enable debugging output



Set to 1 to trace every operation that is performed by mpi-start



Standard input file to use.



Standard output file to use.



Standard error file to use.



Set it to 1 to start only one process per node.



Number of processes to start per node.



Total number of processes to start.

These variables can also be set with the -d command line switch. The following example shows how to set the I2G_MPI_TYPE variable to openmpi:

mpi-start -d I2G_MPI_TYPE=openmpi

There are also other variables that can modify the behaviour of mpi-start, but they are described in other sections of this wiki. The ones dealing with site configuration of mpi-start are documented in the Site Administrator manual, and the variables dealing with the Hooks are summarized in Hooks Framework

Scheduler and Execution Environment Support

mpi-start support different combinations of batch schedulers and execution environments using plugins. The schedulers are automatically detected from the environment and the execution environment can be selected with the I2G_MPI_TYPE variable or the -t command line option.

Scheduler Plugins


supports Grid Engine.


for supporting PBS/Torque.


supports LSF.


gives support for Condor. This plugin lacks the possibility to select how many processes per node should be run.


for supporting Slurm. As with condor, the plugin currently lacks the processes per node support.

Execution Environment Plugins


Open MPI










Debugging environment, just executes application in current host.


The Hooks framework opens the possibility of customizing the behavior of mpi-start. Users can provide their own hooks to perform any pre (e.g. compilation of binaries, data fetching) or post (e.g. storage of application results, clean-up) actions needed for the execution of their application. The Hooks Framework page describes in detail the framework and how to create user hooks.

System configuration

mpi-start can be configured to use the best options for the site. Check Site Administrator manual for more information.


Simple Job

Simple job using environment variables:

   1 #!/bin/sh
   2 # IMPORTANT : This example script execute a
   3 #             non-mpi program with Open MPI
   4 #
   5 export I2G_MPI_APPLICATION=/bin/hostname
   6 export I2G_MPI_TYPE=openmpi

Same example using command line parameters:

mpi-start -t openmpi /bin/hostname

Job with user specified hooks

   1 #!/bin/sh
   2 #
   3 # MPI_START_SHARED_FS can be used to figure out if the current working
   4 # is located on a shared file system or not. (1=yes, 0=no);
   5 #
   6 # The "mpi_start_foreach_host" function takes as parameter the name of
   7 # another function that will be called for each host in the machine as
   8 # first parameter.
   9 # - For each host the callback function will be called exactly once,
  10 #   independent how often the host appears in the machinefile.
  11 # - The callback function will also be called for the local host.
  13 # create the pre-run hook
  14 cat > << EOF
  15 pre_run_hook () {
  16     echo "pre run hook called "
  17     # - download data
  18     # - compile program
  20     if [ "x\$MPI_START_SHARED_FS" = "x0" ] ; then
  21         echo "If we need a shared file system we can return -1 to abort"
  22         # return -1
  23     fi
  25     return 0
  26 }
  27 EOF
  29 # create the post-run hook
  30 cat > << EOF
  31 # the first paramter is the name of a host in the
  32 my_copy () {
  33     CMD="scp . \$1:\$PWD/mydata.1"
  34     echo \$CMD
  35     #\$CMD
  36     # upload data
  37 }
  39 post_run_hook () {
  40     echo "post_run_hook called"
  41     if [ "x\$MPI_START_SHARED_FS" = "x0" ] ; then
  42         echo "gather output from remote hosts"
  43         mpi_start_foreach_host my_copy
  44     fi
  45     return 0
  46 }
  47 EOF
  49 export I2G_MPI_APPLICATION=mpi_sleep
  51 export I2G_MPI_TYPE=openmpi
  52 export I2G_MPI_PRE_RUN_HOOK=./
  53 export I2G_MPI_POST_RUN_HOOK=./
  57 # instead of the variable definition, the following command line could be used:
  58 # mpi-start -t openmpi -pre ./ -post ./ mpi_sleep 0

Using mpi-start with grid middleware


EMI provides the WMS service for submitting jobs to the different available resources. The WMS gets a job description in the JDL language and performs the selection and actual submission of the job into the resources on behalf of the user. The following sections describe how to submit a job using the WMS.

Basic Job Submission

Jobs are described with the JDL language. Most relevant attributes for parallel job submission are:

The following example shows a job that will use 6 processes and it is executed with Open MPI. The requirements attribute makes the WMS to select sites that publish that they support mpi-start and Open MPI.

JobType       = "Normal";
CPUNumber     = 6;
Executable    = "";
Arguments     = "OPENMPI hello_bin hello arguments";
InputSandbox  = {"", "hello_bin"};
OutputSandbox = {"std.out", "std.err"};
StdOutput     = "std.out";
StdError      = "std.err";
Requirements  = member("MPI-START", other.GlueHostApplicationSoftwareRunTimeEnvironment)
                && member("OPENMPI", other.GlueHostApplicationSoftwareRunTimeEnvironment);

The Executable attribute is a script that will invoke mpi-start with the correct options for the execution of the user's application. We propose a generic wrapper that can be used for any application and MPI flavour that gets in the Arguments attribute:

This is the content of the wrapper:

   1 #!/bin/bash
   2 # Pull in the arguments.
   3 MPI_FLAVOR=$1
   5 MPI_FLAVOR_LOWER=`echo $MPI_FLAVOR | tr '[:upper:]' '[:lower:]'`
   8 shift
   9 export I2G_MPI_APPLICATION=$1
  11 shift
  14 # Touch the executable, and make sure it's executable.
  16 chmod +x $I2G_MPI_APPLICATION
  18 # Invoke mpi-start.

User needs to include this wrapper in the InputSandbox of the JDL ( and set it as the Executable of the job. Submission is performed as any other job:

$ glite-wms-job-submit -a

Connecting to the service

====================== glite-wms-job-submit Success ======================

The job has been successfully submitted to the WMProxy
Your job identifier is:


Once the job is finished, the output can be retrieved:

$ glite-wms-job-output

Connecting to the service


                        JOB GET OUTPUT OUTCOME

Output sandbox files for the job:
have been successfully retrieved and stored in the directory:


$ cat /gpfs/csic_projects/grid/tmp/jobOutput/enol_8jG3MUNRm-ol7BqhFP5Crg/std.*
Hello world from gcsic054wn. Process 3 of 6
Hello world from gcsic054wn. Process 1 of 6
Hello world from gcsic054wn. Process 2 of 6
Hello world from gcsic054wn. Process 0 of 6
Hello world from gcsic055wn. Process 4 of 6
Hello world from gcsic055wn. Process 5 of 6

Modifying mpi-start behavior

mpi-start behavior can be customized by setting different environment variables (see usage section for a complete list). If using the generic wrapper, one easy way of customizing mpi-start execution is using the Environment attribute of the JDL. The following JDL adds debugging to the previous example by setting the I2G_MPI_START_VERBOSE and I2G_MPI_START_DEBUG variables to 1:

JobType       = "Normal";
CPUNumber     = 6;
Executable    = "";
Arguments     = "OPENMPI hello_bin hello arguments";
InputSandbox  = {"", "hello_bin"};
OutputSandbox = {"std.out", "std.err"};
StdOutput     = "std.out";
StdError      = "std.err";
Requirements  = member("MPI-START", other.GlueHostApplicationSoftwareRunTimeEnvironment)
                && member("OPENMPI", other.GlueHostApplicationSoftwareRunTimeEnvironment);
Environment   = {"I2G_MPI_START_VERBOSE=1", "I2G_MPI_START_DEBUG=1"};

Use of hooks (see Hooks Framework) is also possible using this mechanism. If the user has a file with the mpi-start hooks called, the following JDL would add it to the execution (notice that the file is also added in the InputSandbox):

JobType       = "Normal";
CPUNumber     = 6;
Executable    = "";
Arguments     = "OPENMPI hello_bin hello arguments";
InputSandbox  = {"", "hello_bin", ""};
OutputSandbox = {"std.out", "std.err"};
StdOutput     = "std.out";
StdError      = "std.err";
Requirements  = member("MPI-START", other.GlueHostApplicationSoftwareRunTimeEnvironment)
                && member("OPENMPI", other.GlueHostApplicationSoftwareRunTimeEnvironment);
Environment   = {"", ""};

eciencia: Middleware/MpiStart/UserDocumentation (last edited 2012-02-17 14:43:17 by enol)