Tree-based FastKick? - Githubissues

amusecode / amuse

Astrophysical Multipurpose Software Environment. This is the main repository for AMUSE

http://www.amusecode.org

Apache License 2.0

158 stars 99 forks source link

Tree-based FastKick? #669

Open rieder opened 4 years ago

rieder commented 4 years ago

FastKick is very useful in Bridged simulations, but it scales with N1*N2, making it still a bottleneck in simulations with many particles (especially where N1 ≈ N2).

Since codes like gadget and phantom don't (yet) support get_gravity_at_point directly, it would be helpful to have a tree-based slim kicker code (preferably OpenMP-enabled) for such simulations. This would scale with log(N1)*N2.

ipelupessy commented 4 years ago

Fi is tree based ad openmp parallel, but I agree that another code might be nice...bonsai on GPU would be nice...

ipelupessy commented 4 years ago

bonsai currently has a host implementation of get_Gravity_at_point...so thats not so nice...maybe something like adding ghost particles with zero mass and then doing a a single gravity evaluation (and removing zero mass parts again) would work..

ipelupessy commented 4 years ago

I am pretty sure @jbedorf should be able to hack something in quickly ;-)

rieder commented 4 years ago

Yes I used Fi for this before, but it seems to complain about a “too short timestep” sometimes... which is weird for a kick-only code, since it doesn’t require a timestep.

ipelupessy commented 4 years ago

yea, it initializes always, necessitating a timestep shift (maybe this could be turned off)

rieder commented 4 years ago

To summarise:

A bug in Bridge was causing codes to do internal force calculations, which prevented them from calculating just the kick
Fi can indeed do this and is now suitable for my purpose
It would still be nice to have a dedicated "TreeKick" code (so I'm keeping this issue open)