Subtask speedup

added enhancement performance scaling labels

@jkeger I have made some more changes to the branch attempting to speed up large gradients. Could you remind me what was going wrong in your case with the last version we tried so it can be addressed?

Brilliant!

I think the main symptom was from the brute-force checks, that a fraction of particles were all missing exactly one neighbour

Can you dig out one setup where this happens?

Sure, I should have that in my notes somewhere for what I was testing.

Actually if I remember right we found the same problems in the included planetary example if you want to try yourself too.

Should I try running one with the latest branch?

That would be helpful if you can tell what setup breaks after how many steps and so on. So we can focus on one issue and eliminate as we go.

Sure, tomorrow's meetings all day so I'll try either early morning or on Friday. So does that mean your recent changes aren't likely to have fixed it?

On the latest subtask_speedup when I make I get configure.ac:112: error: required file 'csds/Makefile.in' not found?

./autogen.sh then ./configure --blabla

Right, running the Planetary/EarthImpact/ example with ./configure --with-hydro=planetary --with-equation-of-state=planetary --with-gravity=basic --with-tbbmalloc --with-parmetis --enable-ipo --enable-hydro-density-checks=100 after a few minutes crashes at ~5.2 ks (shortly after the impact debris starts getting sprayed out), with hydro.c:hydro_exact_density_check():423: N_ngb difference larger than the allowed tolerance for 1 gas particles! (out of 1168 particles)

added 1 commit

• 1de1dc9e - Enable hydro density checks for the planetary scheme

Compare with previous version

added 1 commit

• 5e6d921a - Escape checking whether N_ngb is correct when the particle's h is too close to h_max

Compare with previous version

added 2 commits

• 0a84814f - Do not escape neighbours at the exact same position in DOSELF_SUBSET(). They may genuinely overlap!
• 6f07af07 - Less confusing comments and variable names in hydro's DOSELF_SUBSET()

Compare with previous version

resolved all threads

I have fixed a bug that must have been in the hydro calculation forever, not just in this branch. When a particle needs to redo iterations to fix its h, we were escaping particles at a distance of 0.

Now, for well-behaved hydro schemes that should never happen. Turns out, it does relatively often in this planetary scenario.

I am now able to run the EarthImpact case with checks switched on to completion.

mentioned in issue #687 (closed)

added 1 commit

• 479b944e - Also stop checking whether N_ngb is correct for particles that have reached...

Compare with previous version

added 1 commit

• e60ffcbc - Report on startup that the code is compiled with exact stars density checks

Compare with previous version

I need to close this and re-open via a new branch.

closed