Build with OpenMPI on Mac

[scott-routledge2]

Changes included in this PR

Use OpenMPI for Mac builds to improve initialization times.

• Upgrade mpi4py dependency and remove vendored mpi4py for pip
• Use MPI_Scatterv/MPI_Gatherv for OpenMPI builds (large count APIs aren't supported yet)

Testing strategy

Nightly tests (Run on Linux + openmpi)
https://dev.azure.com/bodo-inc/Bodo/_build/results?buildId=26270&view=results

Release CI

Benchmarking was done on my M2 macbook using the default pixi environment and a simple test program, TPCH SF10, and Taxi (40 million rows)

Results:

Test	MPICH 4.3.2	OpenMPI
test.py	5.3s	2.5s
TPCH SF10 (Total)	196.8s	125.5s
Taxi	10.9s	6.86s

# test.py
import bodo
import bodo.pandas
import time

bodo.set_verbose_level(2)

t0 = time.time()

df = bodo.pandas.DataFrame({'col1': [1, 2, 3], 'col2': [4, 5, 6]})
print("DataFrame created:", df)

print("Total time:", time.time() - t0)

User facing changes

Checklist

• Pipelines passed before requesting review. To run CI you must include [run CI] in your commit message.
• I am familiar with the Contributing Guide
• I have installed + ran pre-commit hooks.

[codecov]

Codecov Report

❌ Patch coverage is 0% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 68.73%. Comparing base (c33fbb5) to head (97a3504).
⚠️ Report is 191 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1010      +/-   ##
==========================================
+ Coverage   66.68%   68.73%   +2.04%     
==========================================
  Files         186      195       +9     
  Lines       66795    67834    +1039     
  Branches     9507     9668     +161     
==========================================
+ Hits        44543    46623    +2080     
+ Misses      19572    18357    -1215     
- Partials     2680     2854     +174     

[scott-routledge2]

mpich should be added to run dependencies automatically (same with impi-devel & openmpi)

[scott-routledge2]

Singleton init (i.e. calling spawn from outside of mpiexec) doesn't seem to be working on github CI runners (maybe a misconfiguration or openmpi bug), it's otherwise been working on the other machines tested so far.

[scott-routledge2]

Just noticed another issue which is when running test.py with the openmpi environment, the get_num_workers function does not behave as expected.

When the program is run outside of mpiexec, MPI.COMM_WORLD.Get_attr(MPI.UNIVERSE_SIZE) is set to 1, so the test.py only spawns 1 worker. When I run with mpiexec, it gets the correct size (10), but I get an error from mpiexec:

>> mpiexec -n 1 python test.py
--------------------------------------------------------------------------
There are not enough slots available in the system to satisfy the 10
slots that were requested by the application:

  env

Either request fewer procs for your application, or make more slots
available for use.

Not sure if this is expected behavior for openmpi or could be the reason CI was behaving strangely.

Would it make sense to ignore the value of MPI.COMM_WORLD.Get_attr(MPI.UNIVERSE_SIZE) (isn't relevant for local macs I think) ?

[IsaacWarren]

Thanks Scott!

[ehsantn]

Thanks @scott-routledge2! Looks good to me.