Skip to content

Latest commit

 

History

History
183 lines (137 loc) · 13.4 KB

README.md

File metadata and controls

183 lines (137 loc) · 13.4 KB
Raw

autotest status Paper Paper Paper Paper Paper Paper

This is Nbody6++GPU - Beijing version, an N-body star cluster simulation code, maintained by Rainer Spurzem ([email protected]) and team.

The code is an offspring of Sverre Aarseth's direct N-body codes see www.sverre.com .

This is the code suitable for parallel and GPU accelerated runs on supercomputers and workstations. Before we give some more practical help, please read the following disambiguation; there is another github of Nbody6++GPU:

LW: https://github.com/nbodyx/ - if interested please contact and collaborate with Long Wang [email protected] RS: https://github.com/nbody6ppgpu - if interested please contact and collaborate with Rainer Spurzem [email protected] [email protected]

Here is an example of current differences between the code version (May 2023), more changes and differences may occur in the future, if in doubt, ask the authors.

  1. LW: implementation of Milky Way potential following the MWPotential2014 in Galpy (Bovy 2015).
  2. RS: implementation of spin and mass dependent recoil kicks after GW merger (Arca Sedda et al. 2023 subm. MNRAS)
  3. LW: implementation of python data reading interface for PeTar analysis tool.
  4. RS: use of HDF5 output files with python data reading interfaces
  5. RS: Namelist based input format, allowing also to read all stellar evolution and binary / collision parameters.
  6. LW and RS: Some bug fixes related to Roche and GR radiation, in both versions slightly different ways.
  7. LW and RS: implementation of BSE from Banerjee et al. 2019

Installation

Get the code

git clone [email protected]:nbody6ppgpu/Nbody6PPGPU-beijing
  1. This downloads the stable branch. The stable branch include major versions, and the dev branch include the most recent updates and bugfix. Changes in dev branch are merged to stable regularly.
  2. If you want the most recent version, use
git clone -b dev [email protected]:nbody6ppgpu/Nbody6PPGPU-beijing

or run git switch dev after you clone without -b dev param.

Configure for compile

./configure [options]
  1. TL;DR: to quickly start on your personal computer, you may use ./configure --enable-mcmodel=large --with-par=b1m --disable-gpu --disable-mpi, and jump to the next section Compile the code
  2. We recommend using --enable-mcmodel=large to allows the program to use much resources.
  3. --with-par=b1m allows up to 1 million particle simulation. In case that your computer has very small memory (<4GB) and your star cluster has a small particle number, you may use smaller value (check ./configure --help for possible value for --with-par)
  4. If you run NBODY6++GPU on your personal computer or workstation rather than computer clusters, MPI can be disabled by append --disable-mpi to the command above.
  5. In the following cases, you may need to append --disable-gpu
  • The computer has no NVIDIA GPU
  • The computer has NVIDIA GPU but did not install CUDA compiler (Test: type nvcc --version in your terminal. If you see information about NVIDIA compiler, then it is installed. If you see errors like "nvcc: command not found" then it is not installed)
  • Your simulation has relatively small particle number (<50000). The code is for up to one million bodies with many initial binaries. In the case of small particle number, GPU can hardly boost the simulation and can sometimes slow it down.
  1. You may set --prefix=[install path] to specify the location to install the executable.
  2. HDF5 is an efficient storage scheme, which is useful during large-scale or long-time simulations to boost the simulation and save disk spaces. Once enabled, the basic particle data (mass, position, velocity) and stellar evolution data will be stored in .h5part files, which may need extra tools to read. HDF5 is recommended but not necessary. You need to install additional libraries to use HDF5. For example, in Debian based Linux sudo apt-get install libhdf5-openmpi-dev libhdf5-dev. After that, append --enable-hdf5 in configure command.
  3. The configure script written by Long Wang has a multitude of further options, check with ./configure --help or feel free to ask any question in our discussion.

Compile the code

make clean
make -j

After make you can find the executable in build/, named nbody6++.[configure-options], where the suffix depends on your configure option (MPI, GPU, HDF5, SIMD, etc), for example nbody6++.avx.mpi.gpu

If you have specified --prefix=[install path] during configure, you may want

make install

and add the installation path to your $PATH environment variable.

Ready for your simulation

  1. (If you have done make install you can skip this step) Copy the executable to the simulation directory you want

    cp `ls build/nbody6++*` [your_simulation_dir]

  2. Prepare an initial condition file. For a test run, you can find example initial conditions in examples/input_files.

    cp examples/input_files/N10k_noDat10.inp [your_simulation_dir]

    This input file let NBODY6++GPU generate a star cluster with 10000 stars with Plummer model, and simulate for only 2 Myr. You can also find N100k.inp and its pre-generated initial particle data dat.10 in examples/input_files for a 100,000 stars, 1 Gyr simulation.

    💡 Starting from the stable version May2023, NBODY6++GPU changes to a fundamentally new and more flexible method of reading input data (control data, not particle data). It uses Fortran NAMELIST input, which has a key=value format. All input data can be given in any order. If you are using a old-format input file, you can use the bash script which transform the old input file into the new one (examples/input_files/@input-transform) to transform it to the new NAMELIST format. See usage inside the script.

  3. CPU and memory

    In simulations with large particle number, segmentation fault may happen. To avoid this, we recommend setting a large OMP_STACKSIZE and disable the memory limitation.

    export OMP_STACKSIZE=4096M
ulimit -s unlimited

    By default, the program uses all CPU threads (which is usually 2 × number of CPU cores). For better performance, OMP_NUM_THREADS should not be too large, and cannot go beyond 32. In case you want to use fewer threads, especially when your computer has more than 32 cores (per node), you need to restrict OMP_NUM_THREADS

    export OMP_NUM_THREADS=[N_threads]

    After running them, you may want to add the these 3 commands to your shell initial file like ~/.bashrc.

  4. Finally, run it

    cd [your_simulation_path]

    If you have done make install and add the installation path to $PATH, run

    nbody6++ < N10k_noDat10.inp

    otherwise you may have copied the executable to the simulation path, run

    ./[your executable filename] < N10k_noDat10.inp

Documentation

To understand the diagnostic information and columns of each output file, please read the documentations at https://www.overleaf.com/read/hcmxcyffjkzq

You are also welcomed to ask any question in our discussion

Data analysis

Some Jupyter notebooks for simple data analysis are provided in examples/. You can check the readme file there to get started.

Tips

  • Before a simulation, it is always recommended to set ulimit -s unlimited before the simulation to avoid segmentation fault.

  • The environment variable OMP_NUM_THREADS has to be set to the desired value of OpenMP threads per MPI process. (Maybe your system has it predefined). I also recommend to set OMP_STACKSIZE=4096M the shell where you run the code.

  • It is inefficient (and even more error prone) for particle numbers below about 50k-100k particles (depending on hardware). For smaller N you are advised to disable GPU, or use Nbody6 and Nbody6GPU for single node/process.

  • It is recommended to provide a dat.10 file in N-body input format (see manual). Such file can be produced by other programs, like McLuster.

Seleted References:

For contributors

git clone -b dev [email protected]:nbody6ppgpu/Nbody6PPGPU-beijing It would automatically switch to dev branch after downloading.

Sources are in src/Main/. Due to urgent bug fixes few routines are later than Dec2020.

Git system does not preserve the modification time of files, but the modification time of some ancient files (created before this project was brought to Git) may be valuable information for developers. If you need this info, run python3 restore_mtime.py after git clone and each git pull. It will touch each file with their real last modification time.

Known Problems:

  1. For systems with more than one GPU on one node the association of MPI rank id and GPU bus id is not well defined, will be improved in next version.

  2. Runs with a million or more bodies and huge numbers of binaries (5% or more) use extreme amounts of computing time for the KS binaries (much much more than should be expected). We work on this.

  3. On some systems heap and stack management when using OpenMP and MPI together seem to produce very strange errors and segmentation faults. The exact reason is not known; we work on this.

  4. Currently using standard OpenMP WITHOUT sse or avx does not work. (it means for configure --disable-simd , but --enable-omp). It uses routines nbint.F instead of special sse or avx routines for neighbour force. We are working on that.

  5. Using much more than one million particles (up to ten million) is still not fully supported. configure already allows --with-par=4m , 8m, 10m, b4m, b8m, b10m . Runs of that size may still fail, depending on your hardware and software environment; also the code may still have some glitches (wrong printout, insufficient vector space allocation); test and work is ongoing.

  6. Many stellar evolution and other parameters are still compiled into the code (see Table A1 in Kamlah et al. 2022, and parameter FctorCl in Rizzuto et al. 2021), mxns0,1 masses of neutron stars; it is the responsibility of the user to keep them all consistent at compile time (for example mxns and FctorCl are defined in two routines independently, see hrplot, coal, mix). We are working to prepare a nice Fortran NAMELIST style input for ALL parameters (the ones from the current input file, and the ones currently compiled in). That will work like in the style of an .ini file with "key=value" pairs and default values.

  7. Currently the use of KZ(7) ge 4 is not working; it produces wrong SIGR2, SIGT2, VROT both in output file and lagr.7. KZ(7) le 3 is ok.

  8. For the latest Aug2022 stable version, there may be problems when using some extreme initial conditions (e.g. multiple massive black holes). We work on this.

Disclaimer

This code and the documentation is given without warranty, hopefully it is helpful. All may contain errors.