sanity check all pickled non-interacting Hamiltonians in h0 directory

[JohanSchott]

• Move sanity checking code of pickled Hamiltonian to a unit-test
• Test all pickled Hamiltonians in h0 directory
• Add sanity checks

[kalvdans]

no need to check events in any particular order.

Suggested change

	for event in process[::-1]:
	for event in process:

[JohanSchott]

The order does not matter in this for-loop. But when a process is used for real, i.e. applied on a product state, the order matters.

The order is chosen such that $c_i^\dagger c_j \ket{\Psi}$ is implemented as process ((i, 'c'), (j, 'a')) operating on the state $\ket{\Psi}$ (with 'c' denoting creation and 'a' annihilation). So the right-most "event" is first applied, and finally the left-most "event", just like the math expression.
Relevant documentation e.g. https://github.com/JohanSchott/impurityModel/blob/master/impurityModel/ed/finite.py#L1320
https://github.com/JohanSchott/impurityModel/blob/master/impurityModel/ed/finite.py#L1395

And in the same way is the Coulomb interaction process between two electrons:
$c_i^\dagger c_j^\dagger c_k c_l$ implemented as ((i, 'c'), (j, 'c'), (k, 'a'), (l, 'a'))
https://github.com/JohanSchott/impurityModel/blob/master/impurityModel/ed/finite.py#L1322

Because of this I think it might be good to consistently loop over events "from right to left" in this repo.

[JohanSchott]

Thanks for your review @kalvdans !

I struggled to understand why the unit-tests failed in this PR.
At first I also thought I could not reproduce the error locally, so I'm sorry about the many git pushes.
But I can reproduce the problem locally, and I have figured out what the problem was.

MPI can't be run both in a parent process and a child process.
In master no unit-test imports MPI stuff, except in a subprocess in impurityModel/test/test_comparison_with_reference.py.
But in this PR add impurityModel/test/test_h0.py which imports MPI stuff with the line from impurityModel.ed.get_spectra import read_pickled_file. So when run pytest in this PR both the parent and a child process uses MPI stuff, which is not ok.

I the last commits I solve this issue by ignoring impurityModel/test/test_comparison_with_reference.py in pytest.ini and running pytest twice:

• pytest impurityModel/test/test_comparison_with_reference.py, which uses MPI only in the child process (and not in parent process).
• pytest, which uses MPI only in the parent process (no subprocess even present in these tests).

Small reproducer of the problem:
python -c "from mpi4py import MPI; import subprocess; output = subprocess.run(args=['mpiexec -n 2 echo hello world'], shell=True, check=False, capture_output=True); print(output)"

But without from mpi4py import MPI it works fine, i.e.:
python -c "import subprocess; output = subprocess.run(args=['mpiexec -n 2 echo hello world'], shell=True, check=False, capture_output=True); print(output)"

[kalvdans]

It seems MPI is automatically initialized when you import mpi4py.MPI (imports should never have side-effects in my opinion).

So just refrain from importing MPI until during runtime when you want to start parallellizing stuff. That also mean that you can't read the constant mpi4py.MPI.COMM_WORLD beforehand :(. And don't test MPI stuff using pytest since you can't restart MPI once you have stopped it...

$ python -c "from mpi4py import MPI; MPI.Finalize(); MPI.Init()"
*** The MPI_Init() function was called after MPI_FINALIZE was invoked.
*** This is disallowed by the MPI standard.
*** Your MPI job will now abort.
[brahe:467487] Local abort after MPI_FINALIZE started completed successfully, but am not able to aggregate error messages, and not able to guarantee that all other processes were killed!

[JohanSchott]

It would be good to test some MPI stuff but not sure how to do that.

I guess one way to test MPI stuff is something like:

def test_MPI_stuff_of_function_XXX():
    tot = []
    for ranks in [1,2,4,8]:
       subproces.run(f"mpiexec -n {ranks} python a_script_that_calls_function_XXX_and_save_its_output.py output_{ranks}", shell=True)
        tot.append(read_file(f"output_{ranks}"))
    compare_data(tot)

where read_file reads e.g. a pickle or hdf5 file.
But it's not so elegant. And one python script for each unit-test...

[kalvdans]

I solve this issue by [...] running pytest twice

Seems it is the same solution pytest-mpi have chosen, and have two decorators to mark up which test should be run in which invocation.

[kalvdans]

Small items left, LGTM!

[kalvdans]

assert_allclose will already print the maximum absolute diff.

[JohanSchott]

true. I'll remove the print statement