On Dec 2, 2013, at 18:22 , Steven R. Brandt sbrandt@cct.lsu.edu wrote:
I'm wondering what it would take to get nonsubcycling AMR.
I think this capability would be very valuable, as it would open up the possibility of doing elliptic solves on each time step (which is something I need to do).
It would also make adaptive time steps easier (something else I need).
Finally, it would help with another problem which I might be able to fix by restructuring.
My update equation computes the second derivative of a variable, and then syncs it so that I have the second derivative everywhere. Even though a sync is called after the derivative is computed, the ghost zones don't get filled in, and even if they did I'm not sure the values that Carpet would compute would make sense (as the data would have to come by prolongation). If we had a nonsubcycling AMR, I could sync whenever I wanted to within an evolution step and not worry about time levels. I may be able to fix this problem by restructuring the code, but I know I need nonsubcycling in the long run anyway.
How hard would it be to teach Carpet how to do this? How hard would it be for me to learn the relevant parts of Carpet to do this?
Steve
It is already possible to use the same time step size on all levels, i.e. to use a time refinement factor of one. This may or may not address your issue.
However, for true uniform time stepping, you would want to avoid subcycling in time. Even with a time refinement factor of one, subcycling in time still introduces differences.
To avoid subcycling in time, you need to rewrite the main evolution loops in the file Evolve.cc. The loops over the refinement levels there need to go away. Instead, these loops need to be present in CallFunction.cc, essentially making global-loop-local the default mode.
Adding a non-subcycling-mode to Carpet would be good. I suggest a telecon to discuss details.
-erik