Skip to main content

CERN code for simulating longitudinal beam dynamics in synchrotrons.

Project description

Pipeline Status Coverage Report Latest Release PyPi Python Documentation Pages

Beam Longitudinal Dynamics Code (BLonD)

Copyright Notice

Copyright 2019 CERN. This software is distributed under the terms of the GNU General Public Licence version 3 (GPL Version 3), copied verbatim in the file LICENCE.txt. In applying this licence, CERN does not waive the privileges and immunities granted to it by virtue of its status as an Intergovernmental Organization or submit itself to any jurisdiction.

Description

CERN code for the simulation of longitudinal beam dynamics in synchrotrons.

Useful Links

Repository: https://gitlab.cern.ch/blond/BLonD

Documentation: https://blond-code.docs.cern.ch/

Project website: http://blond.web.cern.ch

Install

Requirements

  1. A gcc compiler with C++11 support (version greater than 4.8.4).
  2. An Anaconda distribution (Python 3 recommended).
  3. That's all!

Windows GCC Installation Instructions

  1. Download the latest mingw-w64 using this link: https://winlibs.com/#download-release
  2. Extract the downloaded zip/7-zip under e.g. C:\. You should now see a directory C:\mingw64.
  3. Add C:\mingw64\bin in the PATH Environment Variable. Here you can see how to do this in Windows XP/Vista/7/8/10/11.
  4. To validate the correct setup of gcc, open a command prompt and type: gcc --version. The first output line should contain the gcc version you just installed.

Install Steps

Installing BLonD as a python package.

  • Using the pip package manager:

    $ pip install blond
    
  • If this fails try to:

    1. Clone the repository from gitlab or download and extract the zip from https://gitlab.cern.ch/blond/BLonD/-/archive/master/BLonD-master.zip

    2. Go to the downloaded BLonD directory and run:

      $ python setup.py install
      
  • If it still fails, go to the BLonD directory and run:

    1. $ python setup.py compile
      
    2. Then you have to set the PYTHONPATH variable to point to the BLonD installation path.

  • In the occasion that it continues to fail, please open a new gitlab issue.

For advanced users or developers.

  1. You are advised to install git in your system.
  2. Clone the repository or download and extract it.
  3. From within the BLonD directory run:
    $ python blond/compile.py
    
  4. Adjust the PYTHONPATH to contain the path to the cloned repository.

Confirm proper installation

  • Run the unittests with pytest (may need to be installed first with pip install pytest):

    $ pytest -v unittests
    
  • Try to run some of the main files found in the examples:

    $ python __EXAMPLES/main_files/EX_01_Acceleration.py
    $ python __EXAMPLES/main_files/EX_02_Main_long_ps_booster.py
    $ etc..
    

Performace Optimizations

There are some easy ways to reduce the execution time of your simulation:

  1. Use the multi-threaded C library. To use it you have to add the -p flag when compiling the C library:

    $ python blond/compile.py --parallel
    
  2. Enable processor specific compiler optimizations:

    $ python blond/compile.py --flags='-march=native'
    
  3. If you are test-case is calling the synchrotron radiation tracking method, you can accelerate it by using the Boost library. To do so you have to:

    1. Download Boost: https://www.boost.org/. Let's say the version you downloaded is boost_1_70.
    2. Extract it, let's say in /user/path/to/boost_1_70.
    3. Pass the boost installation path when compiling BLonD:
      $ python blond/compile.py --boost=/user/path/to/boost_1_7_70
      
  4. Check the following section about the FFTW3 library.

  5. All the above can be combined.

Changing the floating point precision (32 bit floats or 64 bit floats)

  • By default BLonD uses double precision calculations (float64). To change to single precision, for faster calculations, in the beginning of your mainfile you will have to add the code lines:

    from blond.utils import bmath as bm
    bm.use_precision('single') 
    
  • No other modifications are needed.

Use the FFTW3 library for the FFTs

So far only the rfft(), irfft() and fftfreq() routines are supported. fft_convolve() to be added soon.

  • Windows:

    1. Download and unzip the pre-compiled FFTW3 library. Link: ftp://ftp.fftw.org/pub/fftw/fftw-3.3.5-dll64.zip

    2. Copy the libfftw3-3.dll under your python's distribution directory.

    3. Run the blond/compile.py with the flag --with-fftw.

    4. If the FFTW3 library is not installed in one of the default directories, use the --with-fftw-lib and --with-fftw-header to point to the directories where the shared library and header files are stored.

    5. To use the supported routines, you need to call the function use_fftw() from bmath.py:

      from blond.utils import bmath as bm
      bm.use_fftw()
      ...
      bm.rfft(...)
      bm.irfft(...)
      bm.rfftfreq(...)
      
  • Linux:

    1. Download and compile the FFTW3 library. Link: http://www.fftw.org/fftw-3.3.8.tar.gz

    2. Run the blond/compile.py with the flag: --with-fftw.

    3. If the FFTW3 library is not installed in one of the default directories, use the --with-fftw-lib and --with-fftw-header to point to the directories where the shared library and header files are stored.

    4. Optionally, you can enable one of the flags --with-fftw-omp or --with-fftw-threads to use the multithreaded FFTs.

    5. To use the supported routines, you need to call the function use_fftw() from bmath.py:

      from blond.utils import bmath as bm
      bm.use_fftw()
      ...
      bm.rfft(...)
      bm.irfft(...)
      bm.rfftfreq(...)
      

Using the multi-node (MPI) implementation

Set-up Instructions

  • Add the following lines in your ~/.bashrc, then source it:

    # Environment variables definitions
    export LD_LIBRARY_PATH="$HOME/install/lib"
    export INSTALL_DIR="$HOME/install"
    export BLONDHOME="$HOME/git/BLonD"
    
    # User aliases
    alias mysqueue="squeue -u $USER"
    alias myscancel="scancel -u $USER"
    alias mywatch="watch -n 30 'squeue -u $USER'"
    
    # Module loads
    module load compiler/gcc7
    module load mpi/mvapich2/2.3
    
  • Download and install anaconda3:

    cd ~
    mkdir -p ~/downloads
    cd downloads
    wget https://repo.continuum.io/archive/Anaconda3-2018.12-Linux-x86_64.sh
    bash Anaconda3-2018.12-Linux-x86_64.sh -b -p $HOME/install/anaconda3
    
  • Download and install fftw3 (with the appropriate flags):

    cd ~
    mkdir -p ~/downloads
    cd downloads
    wget http://www.fftw.org/fftw-3.3.10.tar.gz
    tar -xzvf fftw-3.3.10.tar.gz
    cd fftw-3.3.10
    ./configure --prefix=$HOME/install/ --enable-openmp --enable-single --enable-avx --enable-avx2 --enable-fma --with-our-malloc --disable-fortran --enable-shared
    make -j4
    make install
    ./configure --prefix=$HOME/install/ --enable-openmp --enable-avx --enable-avx2 --enable-fma --with-our-malloc --disable-fortran --enable-shared
    make -j4
    make install
    
  • install mpi4py with pip:

    pip install mpi4py
    
  • clone this repo, compile the library and link with fftw3_omp

    cd ~
    mkdir -p git
    cd git
    git clone --branch=master https://github.com/blond-admin/BLonD.git
    cd BLonD
    python blond/compile.py -p --with-fftw --with-fftw-threads --with-fftw-lib=$INSTALL_DIR/lib --with-fftw-header=$INSTALL_DIR/include
    
  • adjust your main file as needed (described bellow).

  • example scripts to setup and run a parameter scan in the HPC Slurm cluster: https://cernbox.cern.ch/index.php/s/shqtotwyn4rm8ws

Changes required in the main file for MPI

  1. This statements in the beginning of the script:

    from blond.utils import bmath as bm
    from blond.utils.mpi_config import worker, mpiprint
    bm.use_mpi()  
    
  2. After having initialized the beam and preferably just before the start of the main loop:

    # This line splits the beam coordinates equally among the workers.
    beam.split()
    
  3. If there is code block that you want it to be executed by a single worker only, you need to surround it with this if condition:

    if worker.isMaster:
        foo()
        ...
    
  4. If you need to re-assemble the whole beam back to the master worker you need to run:

    beam.gather()
    
  5. Finally, in the end of the simulation main loop, you can terminate all workers except from the master with:

    worker.finalize()
    
  6. To run your script, you need to pass it to mpirun or mpiexec. To spawn P MPI processes run:

    $ mpirun -n P python main_file.py
    
  7. For more examples have a look at the __EXAMPLES/mpi_main_files/ directory.

Using the GPU Implementation

Setup Instructions

  1. Install CUDA: https://developer.nvidia.com/cuda-downloads

  2. Install the CuPy library: https://docs.cupy.dev/en/stable/install.html

  3. To verify your installation open a python terminal and execute the following script

    import cupy as cp 
    import numpy as np 
    a = cp.array(np.zeros(1000,np.float64)) 
    

    To compile the Cuda files execute blond/compile.py and add the flag --gpu. The Compute Capability of your GPU will be automatically detected:

$ python blond/compile.py --gpu 

Changes required in the main file for GPU

  1. Right before your main loop you need to add:

    from blond.utils import bmath as bm
    # change some of the basic functions to their GPU equivalent
    bm.use_gpu()
    
  2. Also for every object you are using in your main loop that is in the following list:

GPU objects
Beam
Profile
RingAndRFTracker
TotalInducedVoltage
InducedVoltageTime
InducedVoltageFreq
InductiveImpedance
InducedVoltageResonator
RFStation
BeamFeedback

you need to call their to_gpu() method. The following is a typical example from the __EXAMPLES/gpu_main_files/EX_01_Acceleration.py mainfile.

# Define Objects
beam = Beam(ring, N_p, N_b)
profile = Profile(beam, CutOptions(n_slices=100), 
             FitOptions(fit_option='gaussian'))
# Initialize gpu
beam.to_gpu()
profile.to_gpu()

If an object of this list has a reference inside a different one you don't need to use the to_gpu() for the referenced object. In the previous example, we don't need to call beam.to_gpu() since beam is referenced inside the profile. The same would apply in a TotalInducedVoltage object and the objects in its induced_voltage_list.

Developers

  • Simon Albright (simon.albright (at) cern.ch)
  • Theodoros Argyropoulos (theodoros.argyropoulos (at) cern.ch)
  • Konstantinos Iliakis (konstantinos.iliakis (at) cern.ch)
  • Ivan Karpov (ivan.karpov (at) cern.ch)
  • Alexandre Lasheen (alexandre.lasheen (at) cern.ch)
  • Juan Esteban Muller (JuanF.EstebanMuller (at) ess.eu)
  • Danilo Quartullo (danilo.quartullo (at) cern.ch)
  • Joel Repond (joel (at) repond.ch)
  • Markus Schwarz (markus.schwarz (at) kit.edu)
  • Helga Timko (Helga.Timko (at) cern.ch)

Structure

  • the folder __EXAMPLES contains several main files which show how to use the principal features of the code;
  • the __doc folder contains the source files for the documentation on-line;
  • the various packages which constitute the code are under the blond directory;
  • blond/compile.py is needed to compile all the C++ files present in the project; this file should be run once before launching any simulation. The compiler C++ GCC (at least version 4.8) is necessary.
  • WARNINGS.txt contains useful information related to code usage.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

blond-2.1.2.tar.gz (3.5 MB view hashes)

Uploaded Source

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page