The Taichi MPI demos scaling with MPI4Py.
Anaconda or miniconda is recommended for Python environment management. Official installation guide for miniconda.
- Create the environment and install MPI4Py.
conda create -y -n mpi -c conda-forge python=3.8 mpi4py
- Then activate the environment.
conda activate mpi
- Install Taichi and other dependencies.
pip install -U -r requirements.txt
There are 2 demos in this repository: 00-03 are Poisson solver with 1D and 2D parallelization and 04-06 are LBM solver with 1D parallelization. The LBM code is refactered from the LBM_Taichi repository.
To run:
mpirun -np <num_workers> code.py
All these demos have specific worker numbers, so I just list the run commands below for convenience.
python 00-base-poisson.py
mpirun -n 8 python 01-base-poisson-mpi-band.py
mpirun -n 16 python 02-base-poisson-mpi-halo.py
mpirun -n 16 python 02-base-poisson-mpi-halo.py
mpirun -n 16 python 03-base-poisson-mpi-halo-async.py
python 04-lbm-solver.py
mpirun -np 4 python 05-lbm-solver-mpi.py
mpirun -np 4 python 06-lbm-solver-mpi-faster-border.py
If you cannot reproduce, file an issue in this repo, thanks!
- I haven't tested any MPI4Py programs on supercomputing clusters, it might have problems with different envs. I need a cluster to test these programs.
- Open MPI cannot properly work with the LBM demo, please make sure to use MPICH at the moment. (The conda version install MPICH automatically)
- We are writing a tutorial to explain how to parallelize a stencil computation program written in Taichi with MPI4Py, stay tuned!