Update 3/16

AMR Line Transfer

Buffered sends with a stack of rays is a significant improvement over blocking and looping. Tests on 120 cores:

Buff+Stack Block+Loop Proj
Time ~16 hrs ~30 hrs ~12 hrs
% in LT ~72% ~79% ~50%
Time in LT ~12 hrs ~24 hrs ~6 hrs

I've merged the AMR and projected versions of line transfer into one branch, so it's possible to choose between them with a single switch. Default is non-AMR. During these tests, noticed that AMR version is ~2x faster than non-AMR on 24 cores. Hasn't affected execution time for non-AMR version.

Still probably a way to improve the ray processing order to minimize dependencies, but will take significant thought.

Charge exchange

To get a reasonable time for averaging, need about 9 more frames. Almost there…

https://www.pas.rochester.edu/~adebrech/ChargeExchange/HD209458b_charge_exchange_low0220.png

https://www.pas.rochester.edu/~adebrech/ChargeExchange/HD209458b_charge_exchange_low_species0220.png

MHD

Testing a diffusion (DIFF_ALPHA2) run. Still having pressure protections, but entirely in ghost zones now. Not clear yet if it will improve runtimes if these protections are solved, or if it will behave well (seems like too much diffusion right now).

DIFF_ALPHA = 0

DIFF_ALPHA2 = 0.1

With Diffusion, Frame 39

https://www.pas.rochester.edu/~adebrech/MHD/diff_comp.png

Without Diffusion, Frame 27

https://www.pas.rochester.edu/~adebrech/MHD/MHD_high_res0027.png

Comments

No comments.