- Infos im HLRS Wiki sind nicht rechtsverbindlich und ohne Gewähr -

Difference between revisions of "NEC Cluster Using MPI"

From HLRS Platforms
Jump to navigationJump to search
Line 51: Line 51:
 
of one process per core, no extra threads.
 
of one process per core, no extra threads.
  
 +
For pinning of hybrid OpenMP/MPI, you can use the wrapper from the intel MPI example,
 +
and do not use mpi_paffinity_alone switch, but
  
 +
mpirun -np X -npernode 2 /path/to/wrapper.sh /path/to/app
  
  
Line 93: Line 96:
  
 
With wrapper.sh looking like this
 
With wrapper.sh looking like this
 +
  
 
  #!/bin/bash
 
  #!/bin/bash
 
  export KMP_AFFINITY=verbose,scatter
 
  export KMP_AFFINITY=verbose,scatter
 
  export OMP_NUM_THREADS=4
 
  export OMP_NUM_THREADS=4
  if [ $(expr $PMI_RANK % 2) = 0  ]
+
RANK=${OMPI_COMM_WORLD_RANK:=$PMI_RANK}
 +
  if [ $(expr $RANK % 2) = 0  ]
 
  then
 
  then
        export GOMP_CPU_AFFINITY=0-3
+
      export GOMP_CPU_AFFINITY=0-3
        numactl --preferred=0 --cpunodebind=0 $@
+
      numactl --preferred=0 --cpunodebind=0 $@
 
  else
 
  else
        export GOMP_CPU_AFFINITY=4-7
+
      export GOMP_CPU_AFFINITY=4-7
        numactl --preferred=1 --cpunodebind=1 $@
+
      numactl --preferred=1 --cpunodebind=1 $@
 
  fi
 
  fi
  

Revision as of 16:40, 28 July 2009

OpenMPI example

simple example

To use OpenMPI with intel Compiler, create a .modulerc in your home with this contents:

#%Module1.0#
set version 1.0
module load compiler/intel/11.0
module load mpi/openmpi/1.3-intel-11.0

For compilationuse the mpi wrapper scripts like mpicc/mpic++/mpif90.

The following example is for a pure MPI job, using 16 nodes (128 processes). For Illustration, this is done using an interactvie session (-I option)

First step: Batch submit to get the nodes

 qsub -l nodes=16:nehalem:ppn=8,walltime=6:00:00 -I            # get the 16 nodes

In the session you will get after some time, the application is started with

mpirun -np 128 PathToYourApp

more complex examples

OpenMPI the resources in something called 'slots'. By specifying 'ppn:X' to the batchsystem, the number of slots per node is specified. So for a simple MPI job with 8 process per node (=1 process per core) ppn:8 is best choice, as in above example. Details can be specified on mpirun commandline. PBS setup is adjusted for ppn:8, please do not use other values.

If you want to use less processes per node e.g. because you are restricted by memory per process, or you have a hybrid parallel application using OpenMP and MPI, MPI would always put the first 8 processes on the first node, second 8 on second and so on. To avoid this, you can do

mpirun -np X -npernode 2 /path/to/app

This would start 2 processes per node. Like this, you can use a larger number of nodes with a smaller number of processes, or you can e.g. start threads out of the processes.

If you want to pin your processes to a CPU (and enable NUMA memory affinity) use


mpirun -np X --mca mpi_paffinity_alone 1   /path/to/app

Warning: This will not behave as expected for hybrid multithreaded applications, as the threads will be pinned to a single CPU as well! Use this only in case of one process per core, no extra threads.

For pinning of hybrid OpenMP/MPI, you can use the wrapper from the intel MPI example, and do not use mpi_paffinity_alone switch, but

mpirun -np X -npernode 2 /path/to/wrapper.sh /path/to/app




Intel MPI example

As Nehalem system is a two socket system with local attached ccNUMA memory, memory and process placmeent can be crucial.

Here is an example that shows a 16 node Job, using 1 process per socket and 4 threads per socket and optimum NUMA placement of processes and memory.

Prerequiste: Use intel MPI and best intel compiler To setup environment for this, use this .modulerc file in your home:

#%Module1.0#
set version 1.0
module load compiler/intel/11.0
module load mpi/impi/intel-11.0.074-impi-3.2.0.011

And compile your application using mpicc/mpif90.

First step: Batch submit to get the nodes

 qsub -l nodes=16:nehalem:ppn=8,walltime=6:00:00 -I           # get the 16 nodes

Second step: make a hostlist

 sort -u  $PBS_NODEFILE  > m

Third step: make a process ring to be used by MPI later

mpdboot  -n 16 -f m -r ssh  

Fourth step: start MPI application

mpiexec -perhost 2 -genv I_MPI_PIN 0  -np 32 ./wrapper.sh ./yourGloriousApp

With wrapper.sh looking like this


#!/bin/bash
export KMP_AFFINITY=verbose,scatter
export OMP_NUM_THREADS=4
RANK=${OMPI_COMM_WORLD_RANK:=$PMI_RANK}
if [ $(expr $RANK % 2) = 0  ]
then
      export GOMP_CPU_AFFINITY=0-3
      numactl --preferred=0 --cpunodebind=0 $@
else
      export GOMP_CPU_AFFINITY=4-7
      numactl --preferred=1 --cpunodebind=1 $@
fi


Result is an application running on 16 nodes, using 32 processes spawning 128 threads. One set of 4 therads is pinned to the one socket, the other set of 4 threads to the other socket.