MPI test fails locally

[cjayb]

I've run across this in current master when running test_parallel_backends.py:test_run_mpibackend_oversubscribed, can any other macOS users confirm?

E           RuntimeError: MPI simulation failed. Return code: 143

hnn_core/parallel_backends.py:235: RuntimeError
----------------------------- Captured stdout call -----------------------------
MPI will run over 12 processes
Running 2 trials...
numprocs=12

This is running on my 4-core Air, so I'd expect 1.5 * 4 == 6 cores, not 12. My environment is:

mpi                       1.0                     openmpi    conda-forge
mpi4py                    3.0.3            py38h368b8e5_5    conda-forge
openmpi                   4.1.0                h5807002_1    conda-forge
neuron                    7.8.2                    pypi_0    pypi

[jasmainak]

I have this issue too. Not sure what's the right fix, we might have dig up the pull request by Blake and see what was the intention.

Also, are you sure it's 4 logical cores?

[cjayb]

Also, are you sure it's 4 logical cores?

DUH! No, it's a 1.2 GHz Quad-Core Intel Core i7, so the 12 virtual cores is correct.

[jasmainak]

@cjayb I can't seem to reproduce this error any more. Do you still have this issue? I am now installing mpi4py on Mac by doing the following:

First install openmpi-mpicc:

$ conda install -c conda-forge openmpi-mpicc

Next, do a pip install:

$ pip install mpi4py

The old command:

$ conda install openmpi mpi4py

doesn't seem to work for me any more (conda claims it's installed but pip and ipython don't see it)

[rythorpe]

I still have this issue on Ubuntu. openmpi is installed globally at root but is not installed as a python package.

[cjayb]

Still have the problem with mpi and mpi4py from conda-forge... I need to upgrade to Big Sur soon, so will need to create new envs anyway...

[jasmainak]

moving to 0.3 ... let's wait for some users to hit this issue