Implemented MPI-parallel multilevel diagonal SDC

[brownbaerchen]

Turns out the base transfer class is not compatible with the diagonal sweepers. This is required for PFASSTER.
I don't really know the underlying mathematics, so I just test that the result is the same between the new base_transfer_MPI and base_transfer classes. Let me know if you want anything else tested. There is also a test that a single sweep of multilevel SDC gives the same result with diagonal sweeper and regular sweeper.

I tried pytest-easyMPI and it didn't work for me. Instead, I use subprocess in combination with argparse. I didn't use argparse so far, but will use it a lot from now on. I think it's a very useful library generally, including tests with MPI. Have a look if you're interested in using command line arguments.

[tlunet]

Yep, argparse is really nice indeed (been using it quite a lot already ...).

However, I'm not sure to fully understand why the base transfer class is not compatible with diagonal sweepers ... this should be totally decoupled normally (at least I thought so)

[pancetta]

Yep, argparse is really nice indeed (been using it quite a lot already ...).

However, I'm not sure to fully understand why the base transfer class is not compatible with diagonal sweepers ... this should be totally decoupled normally (at least I thought so)

The transfer classes need to build the tau correction, which in turn needs to calculate QF. And that is distributed in case of diagonal/parallel sweepers. Yes, this is a design flaw, but hey, 10 years ago..

[brownbaerchen]

I just noticed that for some reason this only works with two levels. I'll reopen this PR once I find and fix the bug.

[brownbaerchen]

Bug was in the tau correction which was not computed with only two levels. Test for MPI sweeper now goes up to 3 levels.

[pancetta]

What is the status here?

[brownbaerchen]

Either somebody review it again or merge it, I guess :D

[pancetta]

I kept thinking "Didn't I do this at some point??" and yes, I did:

pySDC/pySDC/projects/parallelSDC/BaseTransfer_MPI.py

Line 20 in a549854

class base_transfer_MPI(object):

[brownbaerchen]

face palm... Looks rather similar :D But are you sure your prolong f works? There's no communication at all in there...
How do you want to proceed? We should have this feature not hidden in a project in the end, but we can use your version if you want.

[pancetta]

As you know, I'm a very good programmer, so, sure, it works 😉 ... that said, can you please remove my code and replace it with yours, making it indeed accessible for "everyone"?

[brownbaerchen]

Using the lovely diff feature of vim, I can see there are some differences after all. Why do you do stuff to the initial conditions in prolong? Eg here. I cannot find something similar in the non MPI version. So is this needed or not? We should have it consistent anyways.

[tlunet]

Using the lovely diff feature of vim, I can see there are some differences after all. Why do you do stuff to the initial conditions in prolong? Eg here. I cannot find something similar in the non MPI version. So is this needed or not? We should have it consistent anyways.

"it might have changed in PFASST" ... wait, what does the might mean ? 😱

[brownbaerchen]

@pancetta, what the hell where you going from when you implemented this MPI version? The stuff in question was removed by you in the non MPI version in 2018, see here :D

[codecov]

Codecov Report

Attention: Patch coverage is 91.76471% with 7 lines in your changes are missing coverage. Please review.

Project coverage is 78.26%. Comparing base (18989d3) to head (931e2e8).
Report is 14 commits behind head on master.

Files	Patch %	Lines
...mentations/sweeper_classes/generic_implicit_MPI.py	25.00%	6 Missing ⚠️
...lementations/sweeper_classes/imex_1st_order_MPI.py	50.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #427      +/-   ##
==========================================
+ Coverage   77.38%   78.26%   +0.87%     
==========================================
  Files         327      330       +3     
  Lines       26085    26201     +116     
==========================================
+ Hits        20187    20507     +320     
+ Misses       5898     5694     -204     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[pancetta]

Relax. "might" means that these things are also called during MLSDC sweeps, but there the u0 is not changed. PFASST however changes the u0 during its iterations and that's why we need to take care of it here. It was not removed by the commit you mention, just done differently.

[brownbaerchen]

Relax. "might" means that these things are also called during MLSDC sweeps, but there the u0 is not changed. PFASST however changes the u0 during its iterations and that's why we need to take care of it here. It was not removed by the commit you mention, just done differently.

Ah, sorry! Was it removed here?

[pancetta]

Hm. Maybe so. In any case, it is currently not there. Tbh, this is all a bit blurry in my head, long time ago. I would assume the base transfer class works as expected and we should go from there.

[brownbaerchen]

I don't really know what's going on. I just took the current version and MPI-ied it. The tests check that MPI and non-MPI versions produce the same results. Let me know if more tests are needed. I don't know if I cover enough cases.

[pancetta]

Well, tests pass, must be correct, no? 😉