topic Distribute processes across nodes in Intel® MPI Library

Distribute processes across nodes

Vishnu — Tue, 04 Jun 2019 14:07:33 GMT

Hi! I'm trying to run a software called LAMMPS, across nodes. As recommended in it's page I'm using 2 OpenMP threads and enough MPI processes to fill all cores: https://lammps.sandia.gov/doc/Speed_intel.html This is fine if I'm running it on one node. When I use, say 4 nodes, the process uses only 2 nodes. How do I distribute it across nodes? This is what I've tried: mpirun -machinefile $PBS_NODEFILE -n 64 -ppn 16 \ -genv OMP_NUM_THREADS=2 -genv I_MPI_PIN_DOMAIN=omp \ lmp -in in.lammps -suffix hybrid intel omp -package intel 0 omp 2 There are 32 cores per node, and so I'm trying to assign 16 MPI processes per node, so each may spawn 2 OMP threads. And `lmp` is the LAMMPS executable. What am I doing wrong?

Hello Vishnu,

Yury_K_Intel — Tue, 04 Jun 2019 15:15:06 GMT

Hello Vishnu,

-machinefile option should not be used together with -n -ppn options. Please replace -machinefile to -f in your command line and try again.

There are 2 ways to set process placement across the nodes:

1. using -machinefile, where you specify the <node_name>:<num_processes> in each line of machine file

2. -f <hostfile> which contains a list of nodes. And using -n you specify a total number of MPI processes and -ppn to specify a number of processes per node.

Best regards, Yury

Hey Yury!

Vishnu — Wed, 05 Jun 2019 04:27:00 GMT

Hey Yury!

Before you replied, I tried this, and it works for me, even with the machinefile:

#!/bin/bash
#PBS -l select=4:ncpus=32:mpiprocs=16
#PBS -N bench
#PBS -q cpuq

NODES=4

cd $PBS_O_WORKDIR

mpirun -machinefile $PBS_NODEFILE -n $((16*${NODES})) -ppn 16 \
    -genv OMP_NUM_THREADS=2 -genv I_MPI_PIN_DOMAIN=omp \
    lmp -in in.lammps -suffix intel -package intel 0 omp 2

The primary difference being that earlier, I was using the following PBS line to requisition nodes:

#PBS -l nodes=4:ppn=32

The application now scales well across nodes.