mpi4py | : Author : Lisandro Dalcin : Contact : dalcinl @ gmail

I'm facing a strange problem when trying to use MPI on a cluster, which uses Slurm as a job scheduler. On both cases, I'm trying to run this simple python program:

I have a C++ class which I intend to call from python's mpi4py interface such that each node spawns the class. On the C++ side, I'm using the Open MPI library (installed via homebrew) and pybind11.

The C++ class is as follows:

I have a function call in python which uses a package method that is non-thread safe (The package writes to three temporary files which have the same name). As the data needed to pass into this method are large and I have numerous input sets, I am approaching this from a distributed perspective, using the MPI4PY library, such that each rank handles a different group of input data at any given time. My problem is that when mapping calls to this function through MPI there are occasions where multiple ranks try to access the function call at once leading to a thread-race condition where data are being overwritten by two calls to the function at once (And then causing the script to error out).

Since the package method is non-thread safe, my question is how would I perform a mutex-style lock on the function such that only one MPI rank is allowed to work inside the function at a time:

For example:

I'm currently trying to understand mpi4py. I set mpi4py.rc.initialize = False and mpi4py.rc.finalize = False because I can't see why we would want initialization and finalization automatically. The default behavior is that MPI.Init() gets called when MPI is being imported. I think the reason for that is because for each rank a instance of the python interpreter is being run and each of those instances will run the whole script but that's just guessing. In the end, I like to have it explicit.

Now this introduced some problems. I have this code

Using mpi4py, I have created a code which defines a new datatype to hold a matrix's column and send it to other MPI process:

I've a big matrix/2D array for which every possible column-pair I need to find the coherence by parallel computation in python (e.g. mpi4py). Coherence [a function] are computed at various child processes and the child process should send the coherence value to the parent process that gather the coherence value as a list. To do this, I've created a small matrix and list of all possible column pairs as follows:

In the Arbor simulator one can specify whether to record no, local, or all spikes when working with distributed MPI simulations. Are there any reasons to record locally on each MPI and broadcasting results versus recording all spikes on just 1 rank with a rank check?

Suppose we have a numpy array.

This link to mpi documentation saves a Julia set as .pgm

Is there any way to alter this code to save the image as .png file?