Redoing AFN neutral limiters #8
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
- This allows other components to pick and choose which ones are used rather than using the total frequency all the time. I am leaving the total frequency in there to be refactored later
Apply limiter/upwinding to both X and Y advection
Seems to work better for neutral perpendicular diffusion than MC.
Take field (Nn, Pn, NVn) and Dnn separately; only upwind the field, and use cell face average values for Dnn. Also add neutral_conduction switch.
target settings used in SOL and PFR regions, rather than SOL and PFR settings.
Seems to help with smoothness of profiles
Same as anomalous_diffusion, obtain flows in X and Y for output when diagnose = true. Note: Only saves perpendicular flux, not parallel flux.
- For debugging purposes
- added setting for pressure floor. Added temporary debug flags for using floored Nn and Pn in DnnNn and DnnPn calculation (otherwise they could unbalance and lead to density loss - need more testing). There is a secondary floor which kicks in at zero gradient, this is now linked to the primary floor (but 2 orders of magnitude below).
- The collision sums contain exactly the same collisions as they did in the past, allowing legacy behaviour. Note this means that collision_frequency != sum(collision_frequencies).
- Makes it easier to split them up later
- Collisionality mode is now a string input to allow different combinations. It will throw an exception if the string is not "legacy" or "braginskii". Still not sure if "legacy" should be called something else or not.
kappa_n is limited upstream and passed down. This seemingly has no impact on performance while limiting conduction fluxes downstream has a huge impact.
Mesh::getBoundaryPar interface changed in BOUT++
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Rewriting the AFN neutral limiter implementation to try and isolate the reason for much worsened performance. The BOUT++ version used for this is here: mikekryjak/BOUT-dev#1
Background
Here is our cross-field neutral transport model:
This pressure-diffusion approach needs flux limitation to a fraction of free streaming to be physical. Our limiters in Hermes-3 are currently crude and unphysical. Here is how they are being changed:
This change introduces a significant slowdown (x10 in small test case to x20 in large production case). This has been isolated to be caused by setting the conduction limiter to be of the new, separate form rather than the previous, Dn derived form.
Note that the new limiters are "smooth" and so should be differentiable. I am using a gamma setting of 4 or 2, which are both very smooth. Gamma is the limiter smoothness (see second slide above). Gamma choice has no impact on the speed.
Initial analysis suggested that freezing
kappaafter the first RHS evaluation, or in the linear solver, can offer significant speedups. This turned out to be an artifact of improperly defined test cases. Here is a comparison of the base (old limiters), new limiters with boolean function, new limiters with asymptotic (smooth) function, and freezing kappa and/or Dn in the linear and total solve:The current effort to debug this is focused around post-processing CVODE residuals. An initial effort with PVODE was tricky because PVODE is ancient and doesn't support preconditioning in the same way, and the slow case ended up faster. The solver also slowed down over time and sped up after a restart, suggesting the solver needs tuning:
