WIP: Do not activate empty stars and BH pairs

Supersedes !1457 (closed) after !1458 (merged) was merged.

Todo:

• Fix the dx_max incompatibility.

assigned to @matthieu

The diff contains chnages that are already merged, like the runner_timestep_collect() function. Is that correct?

No, it isn't. Somehow the merge decided to add that function twice...

added 1 commit

• 856cd6de - Got rid of duplicate runner_do_timestep_collect().

Compare with previous version

added 1 commit

• ea4c019c - Re-added engine_unskip_timestep_communications() prototype to header file.

Compare with previous version

This should compile again now. Git can be so ugly sometimes.

mentioned in merge request !1457 (closed)

approved this merge request

I'll just wait for any flamingo surprised over the next few days but I don't believe there will be any.

Do we want to merge this before we figure out a better way to deal with the drift?

In principle, the drift at the beginning of the step (after the unskip) could reactivate some of the pair tasks that were not unskipped, which would lead to small errors in the neighbour finding which would not happen in the old version. The effect is likely to be small, and since we don't really care about the neighbourhood too much for stars and BH, it is probably not important either. Note that the ghost would not be able to correct for this, since we don't allow the ghost to rerun tasks that were not activated (this would go horribly wrong for domain-crossing pairs).

Ok, I had not understood it was also an issue here and not just in the subtask_speedup branch. So I'll try to think of something over the winter break.

marked this merge request as draft

unapproved this merge request

changed the description

The communication of the time-step fields is likely also necessary in engine_dump_restarts() since we do a drift-all there as well.

Yes, I think you are right. Done.

The big run did once hang in a step that followed a restart dump hence this suggestion. It did restart smoothly from the check-point files.

added 1 commit

• d70dd1bb - Also communicate cell information after drift_all in engine_dump_restarts().

Compare with previous version

resolved all threads

added 487 commits

• d70dd1bb...231cdd06 - 486 commits from branch master
• 1838ec61 - Merge branch 'master' into fewer_timestep_comms_no_empty_pairs

Compare with previous version

Note for self: Merged master in to ease the comparison to !1595 (merged) where the focus is now shifting towards removing unnecessary BH comms too.