Correctly activate the feedback task over the network.

Fixes #561 (closed).

@lhausammann and @aborissov can you confirm that this works?

Unfortunately it doesn't seem to work for me. I tried merging this into my branch and now I don't get the correct evolution even without mpi. Also, non-mpi and mpi evolution are not the same.

mpi:

non-mpi:

The mpi was run with the same setup, only changing the batch script to run with two ranks.

Can you give me the compilation option and runtime flags?

./configure --with-stars=EAGLE --with-chemistry=EAGLE -disable-doxygen-doc --disable-hand-vec --enable-debug=yes --enable-debugging-checks --enable-debug-interactions-stars --enable-optimization=no

I ran the stellar evolution example with ../swift --threads=16 --feedback --external-gravity --stars --hydro stellar_evolution.yml

(for mpi run used swift_mpi and 2 ranks)

I've pushed what I had in my branch to gitlab, there were some changes to the IACT_STARS function calls I had to make in runner_doiact_stars.h because I added more parameters to the eagle feedback functions. I don't think this would impact the changes you made in any way though.

Why did you have to make these changes only now? Weren't the changes already in the branch before?

They were in my branch before. They got overwritten when I merged in master.

added 1 commit

• d705ed96 - Do not forget to activate the self-feedback task in the unskipping MPI case.

Compare with previous version

@aborissov could you try this updated version?

The code seems good to me.

git cherry-pick d705ed96 into your branch is likely the easiest way to test it.

Ok, now I do get the same result between mpi and non-mpi runs. However in both cases it looks like the non-mpi plot above (i.e. the swift result doesn't agree with what we expect based on the standalone test). I'll double check if I'm running something different from what I thought I was running, or if something changed within the feedback functions

Ok, thanks. I think I am activating all the tasks correctly.

The assumptions of the feedback loop are:

• density loop: only update the star properties, do not change the gas
• feedback loop: only update the gas properties, do not change the star.

Note that I have not checked the code against a pure brute-force loop to verify that we actually hit all the neighbours we want.

Could it be the metallicity difference in the parameter file making a come-back?

Ok. Looks like the issue is in the sorted interactions. @aborissov reports that when using --enable-naive-interactions-stars everything is fine. That would mean the issue is not MPI related but rather internal to the PAIR task.

I'll merge this and open another issue to discuss the sorting problem.

merged

mentioned in commit 177e1231

mentioned in issue #563 (closed)

Strange I did try with --enable-naive-interactions-stars and it seemed to be fine

Exactly. That's when it is fine. If you run without it you see problems.