mpi
- Indy Siva
Owned by Indy Siva
MPI users MUST now request mpiprocs in their resource specification
To get 32 cores running across 4 medium nodes
#PBS -l select=4:ncpus=8:NodeType=medium:mpiprocs=8
To get 24 cores running across 4 extra large nodes with only 6 MPI threads per node and without sharing the nodes with other jobs you could use
#PBS -l select=2:ncpus=12:NodeType=xl:mpiprocs=6
Intel MPI Usage
compiler to be used
binaries needs to be compiled using mpi compilers.
We have intel mpi and sgi mpi.
intel mpi
=========
####For intel mpi, use the following syntax####
#module load intel/2019up5/mpi
#mpiexec -n $PROCSÂ $PROGRAM NAME < $INPUT FILE >& $OUTPUT FILE
module load intel-fc-11/12.0.2.137
module load intel-mpi/4.0.1.007
sgi
===
module load mpt/2.02
mpicc mpiexec mpiexec.py mpif90 mpigcc mpiicc mpiifort mpitune mpivars.sh mpicxx mpiexec.hydra mpif77 mpifc mpigxx mpiicpc mpirun.actual mpivars.csh
sample code
cat p1.c
/* Sum processor numbers by passing them around in a ring. */
#include "mpi.h"
#include <stdio.h>
/* Set up communication tags (these can be anything) */
#define to_right 201
#define to_left 102
void main (int argc, char *argv[]) {
int ierror, value, new_value, procnum, numprocs;
int right, left, other, sum, i;
MPI_Status recv_status;
/* Initialize MPI */
MPI_Init(&argc, &argv);
/* Find out this processor number */
MPI_Comm_rank(MPI_COMM_WORLD, &procnum);
/* Find out the number of processors */
MPI_Comm_size(MPI_COMM_WORLD, &numprocs);
/* Compute rank (number) of processors to the left and right */
right = procnum + 1;
if (right == numprocs) right = 0;
left = procnum - 1;
if (left == -1) left = numprocs-1;
sum = 0;
value = procnum;
for( i = 0; i < numprocs; i++) {
/* Send to the right */
MPI_Send(&value, 1, MPI_INT, right, to_right,
MPI_COMM_WORLD);
/* Receive from the left */
MPI_Recv(&new_value, 1, MPI_INT, left, to_right,
MPI_COMM_WORLD, &recv_status);
/* Sum the new value */
sum = sum + new_value;
/* Update the value to be passed */
value = new_value;
/* Print out the partial sums at each step */
printf ("PE %d:\t Partial sum = %d\n", procnum, sum);
}
/* Print out the final result */
if (procnum == 0) {
printf ("Sum of all processor numbers = %d\n", sum);
}
/* Shut down MPI */
MPI_Finalize();
return;
}
compilation
Use mpiicc, mpiifort, or mpiicpc for compilation
module load intel-fc-11/12.0.2.137
module load intel-mpi/4.0.1.007
module load mpi/intel-4.0
mpicc -o p1.out p1.c -lmpi
ls /export/home/snumber/pbs/mpi
mpd.hosts p1.c p1.out
running the job as an mpi job
Manual test run
Please make sure the mpd.hosts file exist. There is a sample mpd.hosts file further down this page . It is only needed for a manual run. It need not be used when the script is run using PBS , either interactively or in batch.
The following syntax will run it on 2 nodes called by the option "-n 2")
mpirun -r ssh -f /export/home/snumber/pbs/mpi/mpd.hosts -n 2 /export/home/snumber/pbs/mpi/p1.out PE 0: Partial sum = 1 PE 0: Partial sum = 1 Sum of all processor numbers = 1 PE 1: Partial sum = 0 PE 1: Partial sum = 1
pbs run
cat pbs.run1
#!/bin/bash -l #PBS -m abe #PBS -V ### Mail to user #PBS -M YOUREMAIL@griffith.edu.au ### Job name #PBS -N mpi_test #PBS -l walltime=01:00:00 ### Number of nodes:Number of CPUs:Number of threads per node #PBS -l select=2:ncpus=4:mpiprocs=4 NPROCS=8 source $HOME/.bashrc ##module load intel-fc-11/12.0.2.137 module load intel-mpi/4.0.0.027 ##module load intel-cc-11/11.1.072 ##module load intel-mpi/4.0.1.007 ##export I_MPI_DEBUG=100 export I_MPI_PLATFORM=auto PBS_EXEC=/opt/pbs/default export I_MPI_MPD_RSH=ssh echo "This jobs runs on the following processors:" echo `cat $PBS_NODEFILE` mpirun -n $NPROCS /export/home/s2594054/pbs/mpi/intelMPI/p1.out 2>&1 echo "Done with job"
qsub pbs.run1 828.pbsserver [snumber@n027 mpi]$ qstat Job id Name User Time Use S Queue ---------------- ---------------- ---------------- -------- - ----- 818.pbsserver 1ivf_apo s2795116 00:00:01 R workq 819.pbsserver 1nn2 s2795116 00:00:01 R workq 821.pbsserver 1ivg s2795116 00:00:00 R workq 825.pbsserver 3nss s2795116 00:00:00 R workq 826.pbsserver 1ivf_naen s2795116 00:00:00 R workq 828.pbsserver mpi s123456 00:00:00 R workq more mpi.o828 Starting job PE 8: Partial sum = 7 PE 9: Partial sum = 8 PE 9: Partial sum = 15 PE 14: Partial sum = 13 PE 3: Partial sum = 2 PE 12: Partial sum = 11 PE 0: Partial sum = 17 PE 16: Partial sum = 15 PE 5: Partial sum = 4 PE 10: Partial sum = 9 PE 10: Partial sum = 17 PE 10: Partial sum = 24 PE 11: Partial sum = 10 <snip>
Important Note
It looks like the wrapper script, mpirun, might not be working properly. mpirun wrapper script actually executes 3 other commands:
mpdboot (to start the MPD daemons), then mpiexec (to run your app), and finally mpdallexit (to close out the MPD ring). You may use the following pbs script as a workaround.
Update: This issue has been resolved by editing the mpdboot.py and pbs_remsh files and hence it is not needed. But keeping the documentation in case it gets broken after any upgrade in the future.
#PBS -m abe
#PBS -M myemail@griffith.edu.au
#PBS -N romsrun-intel
#PBS -l select=3:ncpus=4:mem=4g:mpiprocs=4
#source $HOME/.bashrc
module load intel-fc-11/12.0.2.137
module load intel-mpi/4.0.0.027
module load NetCDF/3.6.3
ROMS_DIR=/export/home/s123456/modelling/roms/forecast/work
HOME_DIR=/export/home/123456
## The number of nodes is given by the select =<NUM > above
NODES=3
#$PBS_NODEFILE is a node-list file created with select and mpiprocs options by PBS
# The number of MPI processes available is mpiprocs * nodes (=NPROCS)
NPROCS=12
export I_MPI_MPD_RSH=ssh
export I_MPI_DEBUG=5
echo "Starting job"
date
${ROMS_DIR}/ocean_eau_forecast_test.in
mpdboot -n $NODES -f $PBS_NODEFILE -r ssh -v
mpdtrace
mpiexec -np $NPROCS ${ROMS_DIR}/oceanG ${ROMS_DIR}/ocean_eau_forecast_test.in
mpdallexit
date
echo "Done with job"
Another intel mpi example
This program calls MPI_Allgather to gather one int from every host:
/*
* * Unit test for Intel MPI failure.
* * mpicc -mt_mpi -o lgcGab lgcGab.c
* */
#include <mpi.h>
#include <stdio.h>
#include <stdlib.h>
int main (int argc, char** argv)
{
int provided;
MPI_Comm comm;
int rank;
int size;
int* sendbuf;
int* recvbuf;
MPI_Init_thread(&argc,&argv,MPI_THREAD_MULTIPLE,&provided);
if (provided<MPI_THREAD_MULTIPLE) {
printf("Need MPI_THREAD_MULTIPLE=%d but got %d\n",
MPI_THREAD_MULTIPLE,provided);
MPI_Abort(MPI_COMM_WORLD,1);
}
MPI_Comm_dup(MPI_COMM_WORLD,&comm);
comm = MPI_COMM_WORLD;
MPI_Comm_rank(comm,&rank);
MPI_Comm_size(comm,&size);
sendbuf = (int*)malloc(sizeof(int));
recvbuf = (int*)malloc(size*sizeof(int));
sendbuf[0] = rank;
//MPI_Gather(sendbuf,1,MPI_INT, recvbuf,1,MPI_INT, 0,comm);
MPI_Allgather(sendbuf,1,MPI_INT, recvbuf,1,MPI_INT, comm);
//MPI_Barrier(comm);
MPI_Finalize();
return 0;
}
mpicc -mt_mpi test.c -o test.o
#!/bin/bash -l
#PBS -m abe
#PBS -V
### Mail to user
####PBS -M YOUREMAIL@griffith.edu.au
### Job name
#PBS -N mpi_test
#PBS -l walltime=100:00:00
### Number of nodes:Number of CPUs:Number of threads per node
#PBS -l select=3:ncpus=2:mpiprocs=2
# This job's working directory
echo Working directory is $PBS_O_WORKDIR
##cd $PBS_O_WORKDIR
source $HOME/.bashrc
module load intel-mpi/4.0.0.027
module load intel-cc-11/11.1.072
echo "Starting job"
echo Running on host `hostname`
echo Time is `date`
echo Directory is `pwd`
export I_MPI_MPD_RSH=ssh
export I_MPI_DEBUG=100
export I_MPI_PLATFORM=auto
PBS_EXEC=/opt/pbs/default
export PATH=${PATH}:${PBS_EXEC}/bin
export PBS_RSHCOMMAND=/usr/bin/ssh
NPROCS=6
#sh /sw/sdev/intel/mpi/4.0.0.027/x86_64/bin/mpivars.sh
#echo This jobs runs on the following processors:
echo `cat $PBS_NODEFILE`
mpirun -n $NPROCS env I_MPI_MPD_RSH=ssh env I_MPI_DEBUG=100 env PATH=$PATH env LD_LIBRARY_PATH=$LD_LIBRARY_PATH /export/home/s123456/pbs/mpi/intelMPI/test.o 2>&1
echo "Done with job"
mpirun -n 6 env I_MPI_MPD_RSH=ssh env I_MPI_DEBUG=100 env PATH=$PATH env LD_LIBRARY_PATH=$LD_LIBRARY_PATH /export/home/s123456/pbs/mpi/intelMPI/test.o
[0] MPI startup(): Intel(R) MPI Library, Version 4.0 Build 20100224
[0] MPI startup(): Copyright (C) 2003-2010 Intel Corporation. All rights reserved.
[0] MPI startup(): Found 1 IB devices
[1] MPI startup(): Found 1 IB devices
[4] MPI startup(): Found 1 IB devices
[5] MPI startup(): Found 1 IB devices
[2] MPI startup(): Found 1 IB devices
[3] MPI startup(): Found 1 IB devices
[1] MPI startup(): Open 0 IB device: mlx4_0
[0] MPI startup(): Open 0 IB device: mlx4_0
[4] MPI startup(): Open 0 IB device: mlx4_0
[5] MPI startup(): Open 0 IB device: mlx4_0
[1] MPI startup(): Start 1 ports per adapter
[2] MPI startup(): Open 0 IB device: mlx4_0
[3] MPI startup(): Open 0 IB device: mlx4_0
[0] MPI startup(): Start 1 ports per adapter
[4] MPI startup(): Start 1 ports per adapter
[5] MPI startup(): Start 1 ports per adapter
[3] MPI startup(): Start 1 ports per adapter
[2] MPI startup(): Start 1 ports per adapter
[0] MPI startup(): shm and ofa data transfer modes
[1] MPI startup(): shm and ofa data transfer modes
[2] MPI startup(): shm and ofa data transfer modes
[3] MPI startup(): shm and ofa data transfer modes
[4] MPI startup(): shm and ofa data transfer modes
[5] MPI startup(): shm and ofa data transfer modes
[0] MPI startup(): I_MPI_DEBUG=100
[0] MPI startup(): I_MPI_FABRICS=shm:ofa
[0] MPI startup(): set domain {0,1,2,3,4,5} fails on node n006
[1] MPI startup(): set domain {0,1,2,3,4,5} fails on node n006
[3] MPI startup(): set domain to {6,7,8,9,10,11,18,19,20,21,22,23} on node n007
[2] MPI startup(): set domain to {0,1,2,3,4,5,12,13,14,15,16,17} on node n007
[5] MPI startup(): set domain to {6,7,8,9,10,11,18,19,20,21,22,23} on node n009
[4] MPI startup(): set domain to {0,1,2,3,4,5,12,13,14,15,16,17} on node n009
[0] MPI startup(): Recognition level=2. Platform code=1. Device=8
[1] MPI startup(): Recognition level=2. Platform code=1. Device=8
[1] MPI startup(): Parent configuration:(intra=1 inter=2 flags=0), (code=1 ppn=1)
[3] MPI startup(): Heterogeneous cluster. Generic defauts are selected
[3] MPI startup(): Recognition level=2. Platform code=1. Device=8
[3] MPI startup(): Parent configuration:(intra=1 inter=2 flags=0), (code=1 ppn=1)
[2] MPI startup(): Heterogeneous cluster. Generic defauts are selected
[2] MPI startup(): Recognition level=2. Platform code=1. Device=8
[2] MPI startup(): Parent configuration:(intra=1 inter=2 flags=0), (code=1 ppn=1)
[0] MPI startup(): Parent configuration:(intra=1 inter=2 flags=0), (code=1 ppn=1)
[0] Allgather: 1: 0-512 & 3-16
[4] MPI startup(): Heterogeneous cluster. Generic defauts are selected
[4] MPI startup(): Recognition level=2. Platform code=1. Device=8
[4] MPI startup(): Parent configuration:(intra=1 inter=2 flags=0), (code=1 ppn=1)
[0] Allgather: 1: 0-2048 & 9-16
[5] MPI startup(): Heterogeneous cluster. Generic defauts are selected
[5] MPI startup(): Recognition level=2. Platform code=1. Device=8
[5] MPI startup(): Parent configuration:(intra=1 inter=2 flags=0), (code=1 ppn=1)
[0] Allgather: 1: 0-512 & 33-2147483647
[0] Allgather: 2: 0-1024 & 17-32
[0] Allgather: 3: 0-2147483647 & 0-2147483647
[0] Allgatherv: 0: 0-2147483647 & 0-2147483647
[0] Allreduce: 1: 0-1024 & 0-2147483647
[0] Allreduce: 1: 0-4096 & 0-4
[0] Allreduce: 1: 0-2048 & 0-8
[0] Allreduce: 1: 0-16384 & 33-2147483647
[0] Allreduce: 2: 1025-2097152 & 3-32
[0] Allreduce: 4: 16385-524288 & 33-2147483647
[0] Allreduce: 3: 0-2147483647 & 0-2147483647
[0] Alltoall: 1: 0-32 & 17-2147483647
[0] Alltoall: 1: 0-128 & 33-2147483647
[0] Alltoall: 2: 0-32768 & 3-32
[0] Alltoall: 2: 32768-2147483647 & 3-4
[0] Alltoall: 3: 0-262144 & 0-2
[0] Alltoall: 3: 513-2147483647 & 33-2147483647
[0] Alltoall: 4: 0-2147483647 & 0-2147483647
[0] Alltoallv: 1: 0-2147483647 & 0-32
[0] Alltoallv: 2: 0-2147483647 & 0-2147483647
[0] Alltoallw: 0: 0-2147483647 & 0-2147483647
[0] Barrier: 2: 0-2147483647 & 0-32
[0] Barrier: 4: 0-2147483647 & 0-2147483647
[0] Bcast: 1: 0-1024 & 0-8
[0] Bcast: 1: 0-8192 & 3-4
[0] Bcast: 7: 0-2147483647 & 0-2147483647
[0] Exscan: 0: 0-2147483647 & 0-2147483647
[0] Gather: 1: 0-64 & 0-2
[0] Gather: 3: 0-2147483647 & 0-2147483647
[0] Gatherv: 1: 0-2147483647 & 0-2147483647
[0] Reduce_scatter: 5: 17-32 & 9-16
[0] Reduce_scatter: 5: 17-64 & 17-32
[0] Reduce_scatter: 5: 0-128 & 33-2147483647
[0] Reduce_scatter: 1: 0-32768 & 3-2147483647
[0] Reduce_scatter: 2: 0-2147483647 & 0-2147483647
[0] Reduce: 3: 524289-2097152 & 0-2
[0] Reduce: 3: 129-8192 & 3-4
[0] Reduce: 1: 0-2147483647 & 0-2147483647
[0] Scan: 0: 0-2147483647 & 0-2147483647
[0] Scatter: 1: 0-2048 & 0-2
[0] Scatter: 3: 0-2147483647 & 0-2147483647
[0] Scatterv: 1: 0-2147483647 & 0-2
[0] Scatterv: 2: 0-2147483647 & 0-2147483647
[0] Rank Pid Node name Pin cpu
[0] 0 12100 n006 n/a
[0] 1 12099 n006 n/a
[0] 2 32239 n007 {0,1,2,3,4,5,12,13,14,15,16,17}
[0] 3 32240 n007 {6,7,8,9,10,11,18,19,20,21,22,23}
[0] 4 3951 n009 {0,1,2,3,4,5,12,13,14,15,16,17}
[0] 5 3952 n009 {6,7,8,9,10,11,18,19,20,21,22,23}
benchmarking intel mpi
module load intel-mpi/4.0.0.027 module load intel-cc-11/11.1.072 NPROCS=4 mpirun -n $NPROCS env I_MPI_MPD_RSH=ssh env I_MPI_DEBUG=5 env PATH=$PATH env LD_LIBRARY_PATH=$LD_LIBRARY_PATH /sw/sdev/intel/mpi/4.0.0.027/x86_64/bin64/IMB-MPI1
openMPI usage
Please note that the syntax for openmpi's mpirun is a little different from intel-mpi's mpirun. Please see a samplle below.
module load mpi/openMPI/1.4.3-gnu ###For openmpi, use the following syntax#### #module load mpi/openmpi/4.0.2 #mpiexec $PROGRAM NAME < $INPUT FILE >& $OUTPUT FILE
Sample run
openMPI has a different mechanism to pass the local environmental variables from master node to slave nodes. You have three options:
1. source ~/.bashrc
2. Use mpirun's --prefix command line option (described below).
3. Modify the wrapper compilers to include directives to include run-time search locations for the Open MPI libraries (see this FAQ entry)
mpirun's --prefix command line option takes as an argument the top-level directory where Open MPI was installed. While relative directory names are possible, they can become ambiguous depending on the job launcher used; using absolute directory names are strongly recommended.
For example, say that Open MPI was installed into /sw/openMPI/1.4.3-gnu/. You would use the --prefix option like this:
shell$ mpirun --prefix /sw/openMPI/1.4.3-gnu/ -np 4 a.out
This will prefix the PATH and LD_LIBRARY_PATH on both the local and remote hosts with /sw/openMPI/1.4.3-gnu/bin and /sw/openMPI/1.4.3-gnu/lib, respectively.
Beginning with the 1.2 series, it is possible to make this the default behavior by passing to configure the flag --enable-mpirun-prefix-by-default. This will make mpirun behave exactly the same as "mpirun --prefix $prefix ...", where $prefix is the value given to --prefix in configure.
Finally, note that specifying the absolute pathname to mpirun is equivalent to using the --prefix argument. For example, the following is equivalent to the above command line that uses --prefix:
/sw/openMPI/1.4.3-gnu/bin/mpirun -np 4 a.out
#PBS -m e
#PBS -M <YourEmail>@griffith.edu.au
#PBS -N Test
#PBS -l select=6:ncpus=4:mem=12g:mpiprocs=4
source $HOME/.bashrc
module load swan/mpi/4085-gnu
## The number of nodes is given by the select =<NUM > above
NODES=6
###$PBS_NODEFILE is a node-list file created with select and mpiprocs options by PBS
###### The number of MPI processes available is mpiprocs * nodes (=NPROCS)
NPROCS=24
echo "Starting job"
/sw/openMPI/1.4.3-gnu/bin/mpirun -machinefile $PBS_NODEFILE -np $NPROCS env PATH=$PATH env LD_LIBRARY_PATH=$LD_LIBRARY_PATH swanrun -input ${SWAN_FORECAST_WORK}/swan_eau_forecast.swn
#
echo "Done with job"
mpd.host file
All compute Nodes
cat mpd.hosts
n001
n002
n003
n004
n005
n006
n007
n008
n009
n010
n011
n012
n013
n014
n015
n016
n017
n018
n019
n020
n021
n022
n023
GPU nodes
cat mpd.hosts
n020
n021
n022
n023
SGI mpi - SGI-mpt
Sample code
cat pingpong.c
#include <stdio.h>
#include "mpi.h"
#define ONE_MIL (1000000)
#define MAX_SIZE_LOG (21)
#define MAX_SIZE (1<<MAX_SIZE_LOG)
char sdata[MAX_SIZE];
char rdata[MAX_SIZE];
main( int argc, char **argv )
{
int rank, i, size, j, next, prev;
long rounds, msglen;
MPI_Status status;
double startt,endt;
char buf[128];
MPI_Init( &argc, &argv );
MPI_Comm_rank( MPI_COMM_WORLD, &rank );
MPI_Comm_size( MPI_COMM_WORLD, &size );
if (rank == 0) {
printf("STARTING LATENCY AND BANDWIDTH BENCHMARK\n");
fflush(stdout);
}
next = (rank + 1 ) % size;
prev = (rank + size - 1 ) % size;
msglen = 0;
while (msglen <= MAX_SIZE) {
if (msglen <= (1<<9)) rounds = 1 << 15;
else if (msglen <= (1<<12)) rounds = 1 << 12;
else if (msglen <= (1<<18)) rounds = 1 << 10;
else rounds = 1 << 8;
MPI_Barrier ( MPI_COMM_WORLD );
startt = MPI_Wtime();
for (j=rounds; j--; ) {
if (rank == 0) {
MPI_Send(sdata,msglen,MPI_BYTE,next,0,MPI_COMM_WORLD);
MPI_Recv(rdata,msglen,MPI_BYTE,prev,0,MPI_COMM_WORLD,&status);
} else {
MPI_Recv(rdata,msglen,MPI_BYTE,prev,0,MPI_COMM_WORLD,&status);
MPI_Send(sdata,msglen,MPI_BYTE,next,0,MPI_COMM_WORLD);
}
}
if (rank == 0) {
double latency,bandwidth, elapse;
endt = MPI_Wtime();
elapse = endt - startt;
latency = elapse / (double)(size*rounds) * (double)ONE_MIL;
bandwidth = ((double)(size*rounds*msglen))/elapse/(double)ONE_MIL;
printf("%3d %8.3f %8.3f\n",(int)msglen,latency,bandwidth);
fflush(stdout);
}
if (msglen == 0) msglen = 1;
else msglen = 2 * msglen;
}
MPI_Finalize();
}
pbs
module avail module load mpt module list Currently Loaded Modulefiles: 1) mpt/2.04
Compile the program
gcc -o pingpong_mpt pingpong.c -lmpi
sample run - masternode
mpirun 2 ./pingpong_mpt cat mpd.hosts n005 n006 mpirun -v n005,n006 2 ./pingpong_mpt mpirun -v n003,n004 2 ./pingpong_mpt MPI: libxmpi.so 'SGI MPT 2.04 03/15/11 05:36:27' MPI: libmpi.so 'SGI MPT 2.04 03/15/11 05:34:18' MPI: libmpi.so 'SGI MPT 2.04 03/15/11 05:34:18' MPI: Cluster collective optimizations enabled. MPI Environmental Settings MPI: MPI_VERBOSE (default: 0) : 1 MPI: Using the InfiniBand interconnect. STARTING LATENCY AND BANDWIDTH BENCHMARK 0 3.770 0.000 1 3.147 0.318 2 3.120 0.641 4 3.126 1.279 8 3.141 2.547 16 3.185 5.024 32 3.213 9.959 64 3.603 17.765 128 3.891 32.896 256 4.243 60.342 512 4.937 103.709
More Examples
hello_mpi
cat hello_mpi.c
/*The Parallel Hello World Program*/
#include <stdio.h>
#include <mpi.h>
main(int argc, char **argv)
{
int node;
MPI_Init(&argc,&argv);
MPI_Comm_rank(MPI_COMM_WORLD, &node);
printf("Hello World from Node %d\n",node);
MPI_Finalize();
}
To compile this:
module load intel-fc-11/12.0.2.137 module load intel-mpi/4.0.1.007 module load mpi/intel-4.0 mpicc -o hello_mpi hello_mpi.c -lmpi
Sample PBS script
cat pbs.run7
#!/bin/bash -l #PBS -m abe ### Mail to user #PBS -M <yourEmail>@griffith.edu.au ### Job name #PBS -N mpi #PBS -l walltime=60:00:00 ## Please note the walltime above . This value * must * be set so that if the ## MPI program runs in an infinite loop , or something similar , it will be ## killed after the given wall time . ### Number of nodes:Number of CPUs:Number of threads per node #PBS -l select=2:ncpus=12:mpiprocs=7 ## The number of nodes is given by the select =<NUM > above NODES=2 ##$PBS_NODEFILE is a node-list file created with select and mpiprocs options by PBS ### The number of MPI processes available is mpiprocs * nodes NPROCS=14 # This job's working directory echo "Working directory is $PBS_O_WORKDIR" cd $PBS_O_WORKDIR source $HOME/.bashrc module load intel-fc-11/12.0.2.137 module load intel-mpi/4.0.1.007 module load mpi/intel-4.0 echo "Starting job" echo Running on host `hostname` echo Time is `date` echo Directory is `pwd` #echo This jobs runs on the following processors: echo `cat $PBS_NODEFILE` mpirun -f $PBS_NODEFILE -n "$NODES" -r ssh -n "$NPROCS" /export/home/snumber/pbs/mpi/2/hello_mpi echo "Done with job"
qsub pbs.run7 2659.pbsserver qstat 2659.pbsserver mpi s123456 00:00:00 R workq cat mpi.o2659 Working directory is /export/home/SNUMBER/pbs/mpi/2 Starting job Running on host n010 Time is Wed Jul 27 08:37:14 EST 2011 Directory is /export/home/SNUMBER/pbs/mpi/2 n010 n010 n010 n010 n010 n010 n010 n020 n020 n020 n020 n020 n020 n020 Hello World from Node 0 Hello World from Node 2 Hello World from Node 4 Hello World from Node 6 Hello World from Node 3 Hello World from Node 1 Hello World from Node 5 Hello World from Node 9 Hello World from Node 11 Hello World from Node 7 Hello World from Node 13 Hello World from Node 8 Hello World from Node 10 Hello World from Node 12 Done with job
qsub -I pbs.run7
cd $PBS_O_WORKDIR
module load intel-fc-11/12.0.2.137
module load intel-mpi/4.0.1.007
module load mpi/intel-4.0
mpirun -f $PBS_NODEFILE -n 2 -r ssh -n 14 /export/home/SNUMBER/pbs/mpi/2/hello_mpi
Hello World from Node 0
Hello World from Node 2
Hello World from Node 1
Hello World from Node 4
Hello World from Node 6
Hello World from Node 3
Hello World from Node 8
Hello World from Node 5
Hello World from Node 13
Hello World from Node 7
Hello World from Node 12
Hello World from Node 10
Hello World from Node 11
Hello World from Node 9
OR
mpirun -r ssh -f $PBS_NODEFILE --totalnum=$NPROCS --verbose -l -machinefile $PBS_NODEFILE -np $(wc -l $PBS_NODEFILE | gawk '{print $1}') /export/home/snumber/pbs/mpi/2/hello_mpi
running mpdallexit on n016
LAUNCHED mpd on n016 via
RUNNING: mpd on n016
LAUNCHED mpd on n020 via n016
RUNNING: mpd on n020
3: Hello World from Node 3
2: Hello World from Node 2
0: Hello World from Node 0
6: Hello World from Node 6
4: Hello World from Node 4
1: Hello World from Node 1
5: Hello World from Node 5
7: Hello World from Node 7
11: Hello World from Node 11
12: Hello World from Node 12
10: Hello World from Node 10
13: Hello World from Node 13
8: Hello World from Node 8
9: Hello World from Node 9
The following could work as well
mpirun -r ssh -f $PBS_NODEFILE --totalnum=$NPROCS --verbose -l -machinefile $PBS_NODEFILE -np 8 /export/home/SNUMBER/pbs/mpi/p1.out
mpirun -r ssh -f $PBS_NODEFILE --totalnum=$NPROCS --verbose -l -machinefile $PBS_NODEFILE -np $(wc -l $PBS_NODEFILE | gawk '{print $1}') /export/home/SNUMBER/pbs/mpi/p1.out
--totalnum specifies the total number of mpds to start
-np number - number of processes
-n <n> or -np <n> # number of processes to start
mpirun -f $PBS_NODEFILE -n $(cat $PBS_NODEFILE | gawk '{print $1}'|sort|uniq|wc -l) -r ssh -n $(wc -l $PBS_NODEFILE | gawk '{print $1}') /export/home/SNUMBER/pbs/mpi/p1.out
To check if mpd is working well:
mpdcheck -f $PBS_NODEFILE -v
PBS select line
#PBS -l select=2:ncpus=8:mpiprocs=8
The line "-l select=2:ncpus=8:mpiprocs=8" is the number of processors required for the mpi job. "select" specifies the number of nodes required; "ncpus" indicates the number of CPUs per node required; and "mpiprocs" represents the number of mpi processes to run per node (normally ncpus=mpiprocs).
As this is not the most intuitive command, the following table is provided as guidance as to how this command works:
select | ncpus | mpiprocs | description |
|---|---|---|---|
4 | 8 | 8 | 32 Processor job, using 4 nodes and 8 processors per node |
4 | 4 | 4 | 16 Processor job, using 4 nodes and 4 processors per node |
16 | 1 | 1 | 16 Processor job, using 16 nodes running 1 mpi process per processor and utilising 1 processor per node |
16 | 8 | 8 | 128 Processor job, using 16 nodes and 8 processors per node (each running an mpi process) |
Ref
1. http://www.hpcu.uq.edu.au/hpc/content/view/225/34/
2. http://www.oerc.ox.ac.uk/computing-resources/osc/support/documentation-help/job-schedulers/pbs/pbs-job-submission-scripts
3. http://www.cardiff.ac.uk/arcca/services/equipment/User-Guide/user-guide.html
PS: tight integration of intelMPI and PBS
For tight integration of intel MPI and PBS, the following needed to be done:
#1 ==> On all execution nodes, add the following:
--> vi /opt/pbs/default/bin/pbs_remsh
while [ $# -gt 1 ]; do
if [ "XX$1" = "XX-j" ]; then
shift;
jobid=$1
shift;
elif [ "XX$1" = "XX-r" ]; then
shift;
rshcmd=$1
shift;
#Add the following 2 lines here.
elif [ "XX$1" = "XX-n" ]; then
shift;
else
break;
fi
done
#2 ==> vi /sw/sdev/intel/mpi/4.0.0.027/x86_64/bin64/mpdboot.py
(search for ssh and find an entry similar to this)
else:
if rshCmd == 'ssh':
#rshArgs = '-x -n -q'
#Added the following entry
rshArgs = '-x -n -q'
elif rshCmd == 'pbs_tmrsh':
rshArgs = ''
elif rshCmd == 'pbs_remsh':
rshArgs = ''
else:
rshArgs = ''
mpdHost = hostsAndInfo[idxToStart]['host']
#3 Sample PBS script
>>>>>>>>>>>>>>>>>>>>>>>
#!/bin/bash -l
#PBS -m abe
#PBS -V
#PBS -M YOUREMAIL@griffith.edu.au
Job name
#PBS -N Intel_mpi_test
#PBS -l walltime=100:00:00
### Number of nodes:Number of CPUs:Number of threads per node
#PBS -l select=2:ncpus=4:mpiprocs=4
NPROCS=8
#cd $PBS_O_WORKDIR
source $HOME/.bashrc
module load intel-mpi/4.0.0.027
module load intel-cc-11/11.1.072
##export I_MPI_MPD_RSH=pbs_remsh
export I_MPI_MPD_RSH=ssh
export I_MPI_DEBUG=100
export I_MPI_PLATFORM=auto
PBS_EXEC=/opt/pbs/default
export PATH=${PATH}:${PBS_EXEC}/bin
mpirun -n $NPROCS /bin/hostname 2>&1
>>>>>>>>>>>>>>>>>>>>>>>
# 4 : Further torubleshooting needed?
If further troubleshooting is needed, add the following entry to /opt/pbs/default/bin/pbs_remsh
-----At the beginning --------
#!/bin/sh
exec >/tmp/remsh.debug 2>&1
set -x
---------------------
At the very bottom, add the following
logger "pbs_remsh cooked options: $remsh $host pbs_attach -j $PBS_JOBID $*"
$remsh "$host" ${PBS_EXEC}/pbs_attach -j "$PBS_JOBID" $*
----------------------
Look at the logs /tmp/remsh.debug and tail -f /var/log/messages to identify issues
Add this entry into : /sw/sdev/intel/mpi/4.0.0.027/x86_64/bin64/mpdboot.py
#!/bin/sh
exec >/tmp/remsh.txt 2>&1
set -x
-------------------
tail /tmp/remsh.txt
n023 env I_MPI_JOB_TAGGED_PORT_OUTPUT=1 HOSTNAME=n022 MPD_CON_EXT=214742.pbsserver_6956 TMPDIR=/scratch/pbs.214742.pbsserver /sw/sdev/intel/mpi/4.0.0.027/x86_64/intel64/bin/mpd.py -h n022 -p 52589 --ifhn=10.110.2.123 --ncpus=1 --myhost=n023 --myip=10.110.2.123 -e -d -s 2
Run it like this to check:
pbs_remsh n023 env I_MPI_JOB_TAGGED_PORT_OUTPUT=1 HOSTNAME=n022 MPD_CON_EXT=214742.pbsserver_6956 TMPDIR=/scratch/pbs.214742.pbsserver /sw/sdev/intel/mpi/4.0.0.027/x86_64/intel64/bin/mpd.py -h n022 -p 52589 --ifhn=10.110.2.123 --ncpus=1 --myhost=n023 --myip=10.110.2.123 -e -d -s 2
tail /tmp/remsh.debug
+ PBS_EXEC=/opt/pbs/default/bin
+ logger 'pbs_remsh cooked options: /usr/bin/ssh -n -n pbs_attach -j 214742.pbsserver n023 env I_MPI_JOB_TAGGED_PORT_OUTPUT=1 HOSTNAME=n022 MPD_CON_EXT=214742.pbsserver_9340 TMPDIR=/scratch/pbs.214742.pbsserver /sw/sdev/intel/mpi/4.0.0.027/x86_64/intel64/bin/mpd.py -h n022 -p 35698 --ifhn=10.110.2.123 --ncpus=1 --myhost=n023 --myip=10.110.2.123 -e -d -s 2'
+ /usr/bin/ssh -n -n /opt/pbs/default/bin/pbs_attach -j 214742.pbsserver n023 env I_MPI_JOB_TAGGED_PORT_OUTPUT=1 HOSTNAME=n022 MPD_CON_EXT=214742.pbsserver_9340 TMPDIR=/scratch/pbs.214742.pbsserver /sw/sdev/intel/mpi/4.0.0.027/x86_64/intel64/bin/mpd.py -h n022 -p 35698 --ifhn=10.110.2.123 --ncpus=1 --myhost=n023 --myip=10.110.2.123 -e -d -s 2
ssh: illegal option -- j
usage: ssh [-1246AaCfgKkMNnqsTtVvXxYy] [-b bind_address] [-c cipher_spec]
[-D [bind_address:]port] [-e escape_char] [-F configfile]
[-i identity_file] [-L [bind_address:]port:host:hostport]
[-l login_name] [-m mac_spec] [-O ctl_cmd] [-o option] [-p port]
[-R [bind_address:]port:host:hostport] [-S ctl_path]
[-w local_tun[:remote_tun]] [user@]hostname [command]
pbs_remsh n023 /bin/hostname
mpirun -r pbs_remsh hostname
mpirun -r pbs_remsh hostnam
mpirun -r pbs_remsh hostname
sh -x /sw/sdev/intel/mpi/4.0.0.027/x86_64/bin64/mpirun -r pbs_remsh hostname
mpdboot -n 2 -f /var/spool/PBS/aux/214742.pbsserver --rsh=pbs_remsh
mpdboot -v -n 2 -f /var/spool/PBS/aux/214742.pbsserver --rsh=pbs_remsh
mpdboot -n 2 -f /var/spool/PBS/aux/214742.pbsserver --rsh=pbs_remsh
sh -x /sw/sdev/intel/mpi/4.0.0.027/x86_64/bin64/mpirun -r pbs_remsh hostname
sh -x /sw/sdev/intel/mpi/4.0.0.027/x86_64/bin64/mpirun -r pbs_remsh hostnameexport MPD_CON_EXT=214742.pbsserver_8221
export MPD_CON_EXT=214742.pbsserver_8221
#5: Try submitting an interactive job
qsub -I -l select=3:ncpus=4:mpiprocs=4 -l walltime=100:00:00
On the execution node:
export I_MPI_MPD_RSH=ssh
export I_MPI_DEBUG=100
NPROCS=12
module load intel-cc-11/11.1.072
module load intel-mpi/4.0.0.027
mpirun -n $NPROCS /bin/hostname
An example from openmpi 4
hello_mpi.c
/*The Parallel Hello World Program*/
#include <stdio.h>
#include <mpi.h>
main(int argc, char **argv)
{
int node;
MPI_Init(&argc,&argv);
MPI_Comm_rank(MPI_COMM_WORLD, &node);
printf("Hello World from Node %d\n",node);
MPI_Finalize();
}
To compile this:
module load mpi/openmpi/4.0.2
mpicc -o hello_mpi hello_mpi.c -lmpi
##mpirun -machinefile $PBS_NODEFILE -np $NPROCS hello_mpi
#!/bin/bash ###PBS -m abe ### Mail to user ##PBS -M <yourEmail>@griffith.edu.au ### Job name #PBS -N mpi ### Number of nodes:Number of CPUs:Number of threads per node #PBS -l select=2:ncpus=2:mpiprocs=2:mem=1g,walltime=10:00:00 ## The number of chunks is given by the select =<NUM > above CHUNKS=2 ##$PBS_NODEFILE is a node-list file created with select and mpiprocs options by PBS ### The number of MPI processes available is mpiprocs * CHUNKS NPROCS=4 # This job's working directory echo "Working directory is $PBS_O_WORKDIR" cd $PBS_O_WORKDIR source $HOME/.bashrc module load mpi/openmpi/4.0.2 echo "Starting job" echo Running on host `hostname` echo Time is `date` echo Directory is `pwd` #echo This jobs runs on the following processors: echo `cat $PBS_NODEFILE` mpirun -machinefile $PBS_NODEFILE env $NPROCS hello_mpi echo "Done with job"