Run MPI On Grid'5000

See also: HPC and HTC tutorial | Run MPI On Grid'5000 | GPUs on Grid5000 | GNU Parallel | Modules | Conda | Guix | Singularity | Deep Learning Frameworks

	Note
	This page is actively maintained by the Grid'5000 team. If you encounter problems, please report them (see the Support page). Additionally, as it is a wiki page, you are free to make minor corrections yourself if needed. If you would like to suggest a more fundamental change, please contact the Grid'5000 team.

Introduction

MPI is a programming interface that enables the communication between processes of a distributed memory system. This tutorial focuses on setting up MPI environments on Grid'5000 and only requires a basic understanding of MPI concepts. For instance, you should know that standard MPI processes live in their own memory space and communicate with other processes by calling library routines to send and receive messages. For a comprehensive tutorials on MPI, see the IDRIS course on MPI. There are several freely-available implementations of MPI, including Open MPI, MPICH2, MPICH, LAM, etc. In this practical session, we focus on the Open MPI implementation.

Before following this tutorial you should already have some basic knowledge of OAR (see the Getting Started tutorial) . For the second part of this tutorial, you should also know the basics about OARGRID (see the Advanced OAR tutorial) and Kadeploy (see the Getting Started tutorial).

Running MPI on Grid'5000

When attempting to run MPI on Grid'5000 you will face a number of challenges, ranging from classical setup problems for MPI software to problems specific to Grid'5000. This practical session aims at driving you through the most common use cases, which are:

• Setting up and starting Open MPI on a default environment using oarsh.
• Setting up and starting Open MPI on a default environment using the allow_classic_ssh option.
• Setting up and starting Open MPI to use high performance interconnect.
• Setting up and starting Open MPI to run on several sites using oargridsub.
• Setting up and starting your own Open MPI library inside a kadeploy image.

Using Open MPI on a default environment

The default Grid'5000 environment provides Open MPI 1.6.5 (see ompi_info).

Creating a sample MPI program

For the purposes of this tutorial, we create a simple MPI program where the MPI process of rank 0 broadcasts an integer (42) to all the other processes. Then, each process prints its rank, the total number of processes and the value he received from the process 0.

In your home directory, create a file ~/mpi/tp.c and copy the source code:

frontend:

mkdir ~/mpi

frontend:

vi ~/mpi/tp.c

#include <stdio.h>
#include <mpi.h>
#include <time.h> /* for the work function only */

int main (int argc, char *argv []) {
       char hostname[257];
       int size, rank;
       int i, pid;
       int bcast_value = 1;

       gethostname(hostname, sizeof hostname);
       MPI_Init(&argc, &argv);
       MPI_Comm_rank(MPI_COMM_WORLD, &rank);
       MPI_Comm_size(MPI_COMM_WORLD, &size);
       if (!rank) {
            bcast_value = 42;
       }
       MPI_Bcast(&bcast_value,1 ,MPI_INT, 0, MPI_COMM_WORLD );
       printf("%s\t- %d - %d - %d\n", hostname, rank, size, bcast_value);
       fflush(stdout);

       MPI_Barrier(MPI_COMM_WORLD);
       MPI_Finalize();
       return 0;
}

You can then compile your code:

frontend:

mpicc ~/mpi/tp.c -o ~/mpi/tp

Setting up and starting Open MPI on a default environment using oarsh

Submit a job:

frontend:

oarsub -I -l nodes=3

oarsh is the remote shell connector of the OAR batch scheduler. It is a wrapper around the ssh command that handles the configuration of the SSH environment. You can connect to the reserved nodes using oarsh from the submission frontal of the cluster or from any node. As Open MPI defaults to using ssh for remote startup of processes, you need to add the option --mca orte_rsh_agent "oarsh" to your mpirun command line. Open MPI will then use oarsh in place of ssh.

	Note
	Debian Wheezy uses plm_rsh_agent instead of orte_rsh_agent

node:

mpirun --mca orte_rsh_agent "oarsh" -machinefile $OAR_NODEFILE ~/mpi/tp

You can also set an environment variable (usually in your .bashrc):

bashrc:

export OMPI_MCA_orte_rsh_agent=oarsh

node:

mpirun -machinefile $OAR_NODEFILE ~/mpi/tp

Open MPI also provides a configuration file for --mca parameters. In your home directory, create a file as ~/.openmpi/mca-params.conf

orte_rsh_agent=oarsh
filem_rsh_agent=oarcp

You should have something like:

helios-52       - 4 - 12 - 42
helios-51       - 0 - 12 - 42
helios-52       - 5 - 12 - 42
helios-51       - 2 - 12 - 42
helios-52       - 6 - 12 - 42
helios-51       - 1 - 12 - 42
helios-51       - 3 - 12 - 42
helios-52       - 7 - 12 - 42
helios-53       - 8 - 12 - 42
helios-53       - 9 - 12 - 42
helios-53       - 10 - 12 - 42
helios-53       - 11 - 12 - 42

You may have (lot's of) warning messages if Open MPI cannot take advantage of any high performance hardware. At this point of the tutorial, this is not important as we will learn how to select clusters with high performance interconnect in greater details below. Error messages might look like this:

[[2616,1],2]: A high-performance Open MPI point-to-point messaging module
was unable to find any relevant network interfaces:

Module: OpenFabrics (openib)
  Host: helios-8.sophia.grid5000.fr

Another transport will be used instead, although this may result in
lower performance.
--------------------------------------------------------------------------
warning:regcache incompatible with malloc
warning:regcache incompatible with malloc
warning:regcache incompatible with malloc

or like this:

[griffon-80.nancy.grid5000.fr:04866] mca: base: component_find: unable to open /usr/lib/openmpi/lib/openmpi/mca_mtl_mx: perhaps a missing symbol, or compiled for a different version of Open MPI? (ignored)
[griffon-80.nancy.grid5000.fr:04866] mca: base: component_find: unable to open /usr/lib/openmpi/lib/openmpi/mca_btl_mx: perhaps a missing symbol, or compiled for a different version of Open MPI? (ignored)
[griffon-80.nancy.grid5000.fr:04865] mca: base: component_find: unable to open /usr/lib/openmpi/lib/openmpi/mca_mtl_mx: perhaps a missing symbol, or compiled for a different version of Open MPI? (ignored)
[griffon-80.nancy.grid5000.fr:04867] mca: base: component_find: unable to open /usr/lib/openmpi/lib/openmpi/mca_mtl_mx: perhaps a missing symbol, or compiled for a different version of Open MPI? (ignored)
...

You could use FAQ#How_to_use_MPI_in_Grid5000.3F to avoid this warnings.

Setting up and starting Open MPI on a default environment using allow_classic_ssh

If you prefer using ssh as a connector instead of oarsh, submit a job with the allow_classic_ssh type:

frontend:

oarsub -I -t allow_classic_ssh -l nodes=3

Launch your parallel job:

node:

mpirun -machinefile $OAR_NODEFILE ~/mpi/tp

	Note
	Using allow_classic_ssh option avoids OAR resources confinement mechanism with cpuset to restrict the jobs on assigned resources. Therefore, allow_classic_ssh cannot be used with jobs sharing nodes between users (i.e. for reservations at the core level).

Setting up and starting Open MPI to use high performance interconnect

By default, Open MPI tries to use any high performance interconnect (e.g. Infiniband) it can find. Options are available to either select or disable an interconnect:

MCA parameters (--mca) can be used to select the drivers that are used at run-time by Open MPI. To learn more about the MCA parameters, see also:

• The Open MPI FAQ about tuning parameters
• How do I tell Open MPI which IP interfaces / networks to use?
• The Open MPI documentation about OpenFabrics (ie: Infiniband)
• The Open MPI documentation about Myrinet

If you want to disable native support for high performance networks, use --mca btl self,sm,tcp. This will switch to TCP backend of Open MPI.

Nodes with Infiniband interfaces also provide an IP over Infiniband interface (these interfaces are named ibX), and can still be used by TCP backend. To also disable their use, use --mca btl_tcp_if_exclude ib0,lo,myri0 or select a specific interface with --mca btl_tcp_if_include eth1. You will ensure that 'regular' Ethernet interface is used.

We will be using NetPIPE to check the performances of high performance interconnects.

To download, extract and compile NetPIPE, do:

frontend:

cd ~/mpi

frontend:

wget http://pkgs.fedoraproject.org/repo/pkgs/NetPIPE/NetPIPE-3.7.1.tar.gz/5f720541387be065afdefc81d438b712/NetPIPE-3.7.1.tar.gz

frontend:

tar -xf NetPIPE-3.7.1.tar.gz

frontend:

cd NetPIPE-3.7.1

frontend:

make mpi

As NetPipe only works between two MPI processes, reserve only one core in two distinct nodes. If your reservation includes more resources, you will have to create a MPI machinefile file with only two entries, such as follow:

frontend:

oarsub -I -l nodes=2

node:

uniq $OAR_NODEFILE | head -n 2 > /tmp/machinefile

Infiniband hardware is available on several sites. For example, you will find clusters with Infiniband interconnect in Rennes (20G), Nancy (20G) and Grenoble (20G & 40G). Myrinet hardware (10G) is available at Lille and Sophia, but is not supported by the current version of operating system used in Grid'5000 (Debian Jessie) (see Hardware page).

To reserve one core of two distinct nodes with:

• a 20G InfiniBand interconnect (DDR, Double Data Rate):

frontend:

oarsub -I -l /nodes=2/core=1 -p "ib20g='YES'"

• a 40G InfiniBand interconnect (QDR, Quad Data Rate):

frontend:

oarsub -I -l /nodes=2/core=1 -p "ib40g='YES'"

To check if the support for InfiniBand is available in Open MPI, run:

node:

ompi_info | grep openib

you should see something like this:

                MCA btl: openib (MCA v2.0, API v2.0, Component v1.4.1)

To start the network benchmark, use:

node:

cd ~/mpi/NetPIPE-3.7.1

node:

mpirun --mca orte_rsh_agent "oarsh" -machinefile $OAR_NODEFILE NPmpi

Without high performance interconnect, results looks like this:

    0:         1 bytes   4080 times -->      0.31 Mbps in      24.40 usec     
    1:         2 bytes   4097 times -->      0.63 Mbps in      24.36 usec     
    ...
    122: 8388608 bytes      3 times -->    896.14 Mbps in   71417.13 usec
    123: 8388611 bytes      3 times -->    896.17 Mbps in   71414.83 usec

The latency is given by the last column for a 1 byte message; the maximum throughput is given by the last line (896.17 Mbps in that case).

With Infiniband 40G (QDR), you should have much better performance that using Ethernet:

 0:       1 bytes  30716 times -->      4.53 Mbps in       1.68 usec
 1:       2 bytes  59389 times -->      9.10 Mbps in       1.68 usec
...
121: 8388605 bytes     17 times -->  25829.13 Mbps in    2477.82 usec
122: 8388608 bytes     20 times -->  25841.35 Mbps in    2476.65 usec
123: 8388611 bytes     20 times -->  25823.40 Mbps in    2478.37 usec

Less than 2 ms of latency and almost 26 Gbit/s of bandwitdh !

More advanced use cases

Running MPI on several sites at once

In this section, we are going to execute a MPI process over several Grid'5000 sites. In this example we will use the following sites: Rennes, Sophia and Grenoble, using oargrid for making the reservation (see the Grid_jobs_management tutorial for more information).

Warning

Open MPI tries to figure out the best network interface to use at run time. However, selected networks are not always "production" Grid'5000 network which is routed between sites. In addition, only TCP implementation will work between sites, as high performance networks are only available from inside a side. To ensure correct network is selected, add the option --mca opal_net_private_ipv4 "192.168.0.0/16" --mca btl_tcp_if_exclude ib0,lo,myri0 --mca btl self,sm,tcp to mpirun

The MPI program must be available on each site you want to use. From the frontend of one site, copy the mpi/ directory to the two other sites. You can do that with rsync. Suppose that you are connected in Sophia and that you want to copy Sophia's mpi/ directoy to Grenoble and Rennes.

fsophia:

rsync -avz ~/mpi/ rennes.grid5000.fr:mpi/

fsophia:

rsync -avz ~/mpi/ grenoble.grid5000.fr:mpi/

(you can also add the --delete option to remove extraneous files from the mpi directory of Rennes and Grenoble).

Reserve nodes in each site from any frontend with oargridsub (you can also add options to reserve nodes from specific clusters if you want to):

frontend:

oargridsub -w 02:00:00 rennes:rdef="nodes=2",grenoble:rdef="nodes=2",sophia:rdef="nodes=2" > oargrid.out

Get the oargrid Id and Job key from the output of oargridsub:

frontend:

export OAR_JOB_KEY_FILE=$(grep "SSH KEY" oargrid.out | cut -f2 -d: | tr -d " ")

frontend:

export OARGRID_JOB_ID=$(grep "Grid reservation id" oargrid.out | cut -f2 -d=)

Get the node list using oargridstat and copy the list to the first node:

frontend:

oargridstat -w -l $OARGRID_JOB_ID | grep -v ^$ > ~/gridnodes

frontend:

oarcp ~/gridnodes $(head -1 ~/gridnodes):

Connect to the first node:

frontend:

oarsh $(head -1 ~/gridnodes)

And run your MPI application:

node:

cd ~/mpi/

node:

mpirun -machinefile ~/gridnodes --mca orte_rsh_agent "oarsh" --mca opal_net_private_ipv4 "192.168.0.0/16" --mca btl_tcp_if_exclude ib0,lo,myri0 --mca btl self,sm,tcp tp

Make your own kadeploy image with latest Open MPI version

Building a kadeploy image

If you need latest Open MPI version, or use Open MPI with specific compilation options (like myrinet/MX support), you must recompile Open MPI from sources. In this section we are going to build an image that includes latest version of Open MPI built from sources. Note that you could also build and install this custom Open MPI in your home directory, without requiring deploying (i.e., using ./configure --prefix=$HOME/openmpi/).

This image will be based on jessie-x64-base. Let's deploy it:

frontend:

oarsub -I -t deploy -l nodes=1,walltime=2

frontend:

kadeploy3 -f $OAR_NODEFILE -e jessie-x64-base -k

Connect to reserved node as root:

frontend:

ssh root@$(head -1 $OAR_NODEFILE)

Download Open MPI sources from official website:

frontend:

cd /tmp/

frontend:

wget http://www.open-mpi.org/software/ompi/v1.10/downloads/openmpi-1.10.2.tar.bz2

node:

tar -xf openmpi-1.10.2.tar.bz2

node:

cd openmpi-1.10.2

Install build dependencies:

node:

apt-get -y install g++ make gfortran f2c libblas-dev

Configure, compile and install:

node:

./configure --libdir=/usr/local/lib64 --with-memory-manager=none

node:

make -j8

node:

make install

To run our MPI applications, we create a dedicated user named mpi. We add it to the rdma group to allow access to Infiniband hardware. Also, we copy the ~root/authorized_keys files so that we can login as user mpi from the frontend. We also create an SSH key for identifying the mpi user (needed by Open MPI).

useradd -m -g rdma mpi -d /var/mpi
echo "* hard memlock unlimited" >> /etc/security/limits.conf
echo "* soft memlock unlimited" >> /etc/security/limits.conf

mkdir ~mpi/.ssh
cp ~root/.ssh/authorized_keys ~mpi/.ssh
chown -R mpi ~mpi/.ssh
su - mpi
ssh-keygen -N "" -P "" -f /var/mpi/.ssh/id_rsa
cat .ssh/id_rsa.pub >> ~/.ssh/authorized_keys
echo "        StrictHostKeyChecking no" >> ~/.ssh/config
exit # exit session as MPI user
exit # exit the root connection to the node

# You can then copy your file from the frontend to the mpi home directory:
rsync -avz ~/mpi/ mpi@$(head -1 $OAR_NODEFILE):mpi/ # copy the tutorial

Save the newly created image by using tgz-g5k:

node:

ssh root@$(head -1 $OAR_NODEFILE) tgz-g5k > $HOME/public/jessie-openmpi.tgz

Copy the description file of jessie-x64-base:

frontend:

kaenv3 -p jessie-x64-base | grep -v visibility > $HOME/public/jessie-openmpi.dsc

Change the image name in the description file; we will use an http URL for multi-site deployment:

frontend:

sed -i "s,server://.*images/jessie-x64-base.*,http://public.$(hostname | cut -d. -f2).grid5000.fr/~$USER/jessie-openmpi.tgz," $HOME/public/jessie-openmpi.dsc

Release your job:

frontend:

oardel $OAR_JOB_ID

Using your Kadeploy image

Single site

Reserve some nodes and deploy them:

frontend:

oarsub -I -t deploy -l /nodes=3

frontend:

kadeploy3 -a $HOME/public/jessie-openmpi.dsc -f $OAR_NODEFILE -k

Copy machines file and connect to first node:

frontend:

scp $OAR_NODEFILE mpi@$(head -1 $OAR_NODEFILE):nodes

frontend:

ssh mpi@$(head -1 $OAR_NODEFILE)

Copy your MPI application to other nodes and run it:

node:

cd ~/mpi/

node:

/usr/local/bin/mpicc tp.c -o tp

node:

for node in $(uniq ~/nodes | grep -v $(hostname)); do scp ~/mpi/tp $node:~/mpi/tp; done

node:

/usr/local/bin/mpirun -machinefile ~/nodes ./tp

Multiple sites

Choose three clusters from 3 different sites.

frontend:

oargridsub -t deploy -w 02:00:00 cluster1:rdef="nodes=2",cluster2:rdef="nodes=2",cluster3:rdef="nodes=2" > oargrid.out

frontend:

export OARGRID_JOB_ID=$(grep "Grid reservation id" oargrid.out | cut -f2 -d=)

Get the node list using oargridstat:

frontend:

oargridstat -w -l $OARGRID_JOB_ID | grep grid > ~/gridnodes

Deploy on all sites using the --multi-server option :

frontend:

kadeploy3 -f gridnodes -a $HOME/public/jessie-openmpi.dsc -k --multi-server

Copy machines file and connect to first node:

frontend:

scp ~/gridnodes mpi@$(head -1 ~/gridnodes):

frontend:

ssh mpi@$(head -1 ~/gridnodes)

Copy your MPI application to other nodes and run it:

node:

cd ~/mpi/

node:

/usr/local/bin/mpicc tp.c -o tp

node:

for node in $(uniq ~/gridnodes | grep -v $(hostname)); do scp ~/mpi/tp $node:~/mpi/tp; done

node:

/usr/local/bin/mpirun -machinefile ~/gridnodes --mca btl self,sm,tcp --mca opal_net_private_ipv4 "192.168.0.0/16" --mca btl_tcp_if_exclude ib0,lo,myri0 --mca orte_keep_fqdn_hostnames 1 tp

See also: HPC and HTC tutorial | Run MPI On Grid'5000 | GPUs on Grid5000 | GNU Parallel | Modules | Conda | Guix | Singularity | Deep Learning Frameworks