Using MPI in Python (mpi4py)¶
Intro to mpi4py and an example on Cori¶
mpi4py provides MPI standard bindings to the Python programming language. Documentation on mpi4py is available here.
Here is an example of how to use mpi4py on Cori using Anaconda Python 3.7. Consider this minimal example program:
1 2 3 4 5
This program will initialize MPI, find each MPI task's rank in the global communicator, find the total number of ranks in the global communicator, print out these two results, and exit. Finalizing MPI with mpi4py is not necessary; it happens automatically when the program exits.
Suppose we put this program into a file called "mympi.py." To run it on the Haswell nodes on Cori, we could create the following batch script in the same directory as our Python script, that we call "myjob.sh:"
1 2 3 4 5 6 7
To run "mympi.py" in batch on Cori, we submit the batch script from the command line using sbatch, and wait for it to run:
% sbatch myjob.sh Submitted batch job 987654321
After the job finishes, the output will be found in the file "slurm-987654321.out:"
% cat slurm-987654321.out ... 91 96 44 96 31 96 ... 0 96 ...
mpi4py in your custom conda environment¶
If you would like to use mpi4py in a custom conda environment, you will need to install and build it inside your environment.
Do NOT conda/pip install mpi4py
You can install mpi4py using these tools without any warnings, but your mpi4py programs just won't work. To use Cori's MPICH MPI, you'll need to build it yourself using the Cray compiler wrappers that link in Cray MPICH libraries.
You can build
mpi4py and install it into a conda environment on Cori using a recipe like the following:
wget https://bitbucket.org/mpi4py/mpi4py/downloads/mpi4py-3.0.3.tar.gz tar zxvf mpi4py-3.0.3.tar.gz cd mpi4py-3.0.3 module swap PrgEnv-intel PrgEnv-gnu python setup.py build --mpicc="$(which cc) -shared" python setup.py install
Bug in conda-provided
ld has been resolved
The December 5, 2019 Cori maintenance exposed a bug in Anaconda's compiler compatilibity
ld GNU linker. This issue was resolved on Jan 10, 2020.
The MPI-enabled Python interpreter is not required (see this page in the mpi4py documentation). To install it however, use these additional steps:
python setup.py build_exe --mpicc="$(which cc) -dynamic" python setup.py install_exe
MPI_COMM_WORLD size is 1 ?!?!
If you try to use mpi4py and you observe something like an apparent
MPI_COMM_WORLD size of 1 and all processes report that they are rank 0, check to see if you have installed mpi4py from Anaconda with the Conda tool (which will not work on our systems). If you have, scroll back up and see the directions about how to build mpi4py correctly in your conda environment.
Ok so now you have mpi4py built and ready to use. Make sure you grab a compute node either via the interactive queue or with sbatch. MPI is disabled on our login nodes to prevent users from running their expensive computations there. If you try to use MPI on a login node you'll see this warning:
MPI doesn't work on NERSC login nodes
Initializing MPI on a login node will not work at NERSC. This is what you will see if you try to do it:
nersc$ module load python nersc$ python -c 'from mpi4py import MPI' [Fri Aug 9 09:26:55 2019] [unknown] Fatal error in PMPI_Init_thread: Other MPI error, error stack: MPIR_Init_thread(537): MPID_Init(246).......: channel initialization failed MPID_Init(647).......: PMI2 init failed: 1 Aborted
If you see this kind of output from a batch job or in an interactive allocation then it means something different. It likely means that
MPI_Init() exceeded a timeout, perhaps due to I/O issues. This is more likely to occur when the file system you are importing packages from isn't optimized for serving up code to the compute nodes. Increasing the timeout is a temporary fix:
/global/common/softwareor better yet, Shifter.
About Huge Memory Pages (Updated 2020-03-20)
Previously, from 2019-08-02 until this 2020-03-20, we had recommended users unload craype-hugepages2M before compiling due to an issue with Python and huge memory pages. Cray fixed the issue, so we have removed the guidance to unload that particular module. Please let us know if you see any problems arise due to this change.
Using mpi4py in Shifter containers¶
If you are planning to use mpi4py, you should also seriously consider using Shifter.
This is more work in the short term but will pay off in the long term. Your startup times will be lightning fast and and you will not be at the mercy of whatever jobs may also be running at NERSC (and slowing down the filesystem).
For more information about how to build and use mpi4py in a Shifter container, please see this page.