CEMRACS 2017

Numerical methods for stochastic models: control, uncertainty quantification, mean-field

July 17 - August 25, CIRM, Marseille

[ Be prepared | Mésocentre | Lectures | Training sessions ]

High Performance Computing

Be prepared for the training sessions

You will need your own laptop to write and test the code implemented during the training sessions. Once tested locally, you will able to run your codes on high performance computing plaforms kindly put at CEMRACS' disposal by Aix-Marseille University mésocentre.
On your a laptop, you need to have

a C/C++ compiler with OpenMP support;
the OpenMPI library;
CMake;
the PNL library.

You will find below some explanations on how to get it done on Linux and Mac OS X systems.

Linux

Use the package manager of your distribution to install

A C compiler with OpenMP support: either gcc/g++ or Clang ≥ 3.7 (the package may be called LLVM).
The openMPI library. This library may be split in two parts depending on your distribution: openmpi-bin and openmpi-dev. Note that the library may be called libopenmpi rather than openmpi.
CMake >= 3.0.

For the installation of PNL see below.

Mac OS X

First, you need to install a recent version of XCode from the Apple Store. Then, open a Terminal and run

xcode-select --install

to install the Command Line Tools.

To get the other functionalities, we recommend you to use the Homebrew package manager. If you already have MacPorts, skip this part and go to the MacPorts section.

Homebrew

Install Homebrew as explained on Homebrew homepage.

To install the GNU compiler collection (gcc/g++), run
```
brew install gcc
```
As of now, it installs version 7 of the GNU compiler collection. The C compiler is gcc-7 and the C++ compiler is g++-7. They both have support for openMP.
To install openMPI, run
```
brew install openmpi
```
To install CMake, run
```
brew install cmake
```

For the installation of PNL see below.

MacPorts

Install MacPorts as explained on MacPorts homepage.

To know the currently available version of the specified ports, use the command
```
port search --name 'gcc*' 
```
To install the gcc compiler infrastructure, run
```
port install gcc7 (or 6)
```
It provides a C compiler and a C++ compiler with openMP support.
To install openMPI, run
```
port install openmpi
```

For the installation of PNL see below.

Installing PNL on Mac OS X or Linux

The PNL library is hosted on github: https://github.com/pnlnum/pnl.
To get the library, you can either clone the Git repository

git clone https://github.com/jlelong/pnl.git

or directly download a zip file: pnl-master.zip. Unzipping the file will create a pnl-master repository. To compile the library, just run the following commands

mkdir pnl-build
cd pnl-build
cmake -DCMAKE_BUILD_TYPE=Debug /relative/path/to/pnl
make
make install

If you installed a specific compiler to have the OpenMP support, you need to compile PNL with it by running CMake as

 cmake -DCMAKE_BUILD_TYPE=Debug -DCMAKE_C_COMPILER=<name_of_C_compiler> -DCMAKE_CXX_COMPILER=<name_of_C++_compiler> /relative/path/to/buil-dir

Working on the machines of the mesocentre

Front-end connection:

$ ssh user@login.ccamu.u-3mrs.fr
user@login's password:

To change your password, use the command passwd.
You will need a password free SSH access using SSH keys. First, generate a pair of keys or jump to the next step if you already have SSH keys

ssh-keygen -t rsa

Pick up the passphrase of your choice. Then, copy the public key on the front-end

ssh-copy-id user@login.ccamu.u-3mrs.fr

Development environment

The environment modules is used to define libraries and/or path to compilers.

available modules: module avail
the compiler gcc 4.7.3 can be loaded with: module load gcc/4.7.3
The loaded modules are available with: module list
To unload a module: module unload gcc/4.7.3

The following modules will be used for the training sessions: openmpi/gcc47/1.6.3, LAPACK/gcc47/3.4.0, ATLAS/gcc47/3.8.4, gcc/4.7.3, cmake/3.5.2.

Environment for MPI on the cluster of the mesocentre

To write and execute MPI's program, load de modules:

module load gcc/4.7.3 
module load openmpi/gcc47/1.6.3

Job scheduler: OAR

To submit jobs, you must use the job scheduler OAR.
Basic commands are:

oarsub -I: interactive submission;
oarsub -S ./mon_script.oar : OAR script submission;
oarstat: show the submitted jobs (see also Monika);
oarsub -C JOB_ID : to connect to compute nodes being used. The JOB_ID can be obtained with oarstat.
On the master node, it is possible to connect to other nodes by achieving the command oarsh name_node where name_node can be obtained with: cat $OAR_NODEFILE;
oardel JOB_ID: delete the job. The JOB_ID can be obtained with oarstat.