Repository navigation
IBM: ib_flux_wall for fixed no-slip inert bodies, and trace species no longer starve the reacting flux wall - #2213
Merged
sbryngelson merged 4 commits intoNov 14, 2026
Conversation
The positivity factor theta of a fluid cell caps all of its wall fluxes so that they take at most 0.9 of any consumed species in one forward-Euler step. A species with a negligible share of the consumption that the cell holds almost none of (OH: rho Y ~ 1e-46 against a wall rate ~ 1e-45 per step) set theta ~ 0 and starved the whole wall. Species carrying less than 1e-6 of the cell's wall consumption no longer set theta. On a 48 x 48 grid the reacting-cylinder test starved from step ~27 and the burn rate fell 12 % by 2e-5 s; the 24 x 24 Flux Wall test is unaffected (theta stays 1), so its golden does not change. Adds the 48 x 48 case as a test and updates the flux-wall columns of the mass-budget example table.
ib_flux_wall now also covers fixed, no-slip, non-blowing inert bodies (surface_reaction = 0): their wall faces keep a zero source, so they carry no mass or species, only conduction, pressure and viscous stresses. This removes the ghost-cell gas exchange across the staircase faces that makes the gas mass of a closed periodic box with fixed isothermal inert bodies drift. An inert body that slips or blows keeps the ghost-state wall. The option stays opt-in (default F); results with it off do not change. The case validator accepts the option with such inert bodies (it still needs at least one reacting or no-slip, non-blowing body, and every body fixed). Adds two inert flux-wall tests: the reacting cylinder example with surface_reaction = 0, and a closed periodic box.
|
Claude Code Review Head SHA: 1b0d007 Files changed:
Findings: |
f_wall_face_sum built per-thread arrays h(3) and c(3), which the amdflang OpenMP-offload build placed in scratch memory. It now sums the neighbour terms through f_wall_face_term, so only the order of the floating-point additions changes. MI300A (amdflang AFAR, 96 cells/D, 100 steps, rocprofv3): s_ib_flux_wall_prepare took 3.69 ms per call (1107 ms of the 1950 ms of kernel time) before and 0.030 ms per call (9.0 ms in total) after. ib_flux_wall step cost went from +80% (9.6 -> 17.4 ns/gp/eq/rhs) to within run-to-run noise (10.7 on vs 9.7 and 11.6 off).
Lines of Code
|
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## master #2213 +/- ##
==========================================
+ Coverage 65.15% 65.16% +0.01%
==========================================
Files 110 110
Lines 29555 29559 +4
Branches 4476 4476
==========================================
+ Hits 19256 19263 +7
+ Misses 6743 6742 -1
+ Partials 3556 3554 -2 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
sbryngelson
added a commit
to sbryngelson/MFC
that referenced
this pull request
Nov 14, 2026
…s ghost temperature also moves 8B2A7F3E, FAF04D7A and 3F5FE078 (MFlowCode#2213 inert and trace-species flux walls) and F8D466BD and C02C399E (MFlowCode#2200 thermostat, goldens regenerated for MFlowCode#2205) were generated on master without the isothermal ghost change.
This file contains hidden or 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
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.
Two commits on
origin/master80997da, each with its own tests and goldens. The option stays opt-in (ib_flux_walldefaultF); with it off nothing changes (no existing golden moved on either vendor, see Tests).Summary
m_ibm.fpp,s_ib_flux_wall_prepare). The positivity factorthetaof a fluid cell caps all of its wall faces so that they take at most 0.9 of any consumed species in one forward-Euler step. A species with a negligible share of the wall consumption that the cell holds almost none of settheta~ 0 for every species. Species carrying less than 1e-6 of the cell's wall consumption no longer settheta(they are not limited;rhoYof such a species can in principle go slightly negative, which no run here showed).m_ibm.fpp,s_ib_flux_wall_setup, validator, docs).ib_flux_wallalso covers fixed, no-slip, non-blowing bodies withsurface_reaction = 0. Their wall faces keep a zero source: no mass or species, only conduction, pressure and viscous stresses. This removes the ghost-cell gas exchange across the staircase faces that makes the gas mass of a closed periodic box with fixed isothermal inert bodies drift. An inert body that slips or blows keeps the ghost-state wall. The validator accepts the option with such bodies (it still needs every body fixed and at least one reacting or no-slip, non-blowing body). In a case that mixes reacting and inert bodies withib_flux_wall = T, the inert bodies now get the zero-flux wall.No case sets
ib_flux_wallby default, and no toolchain default is touched.Evidence
Machines: A100: NVIDIA A100 80GB PCIe, nvfortran 25.11, OpenACC (
--gpu acc), 1 rank unless stated. MI300A: AMD MI300A (gfx942), amdflang (AFAR 24.3.0), OpenMP offload (--gpu mp), 1 rank. "master" is a build oforigin/master80997da; "branch" is this PR. The MI300A runs below used the branch with the option off, which is bit-identical to master (see Tests); they reproduce the A100 master numbers to all printed digits.1. Trace species (commit 1)
Mass-budget example (
examples/2D_ibm_wall_mass_budget, reacting carbon cylinder in a closed box, 16 cells/D,--flux_wall, A100). Gasified mass is the time integral of the surface records, as inmass_budget.py; the fix is the only difference.The gas budget closes to 0.3% or better after 60 us in both. That the shortfall on master is trace-species starvation is inferred from the next table; I did not print
thetafor this case.The 48 x 48 reacting-cylinder case (the new test; the 24 x 24
Flux Walltest never hastheta< 1). Gasification rate of the body (ib_heat_wrtcolumn 3), A100:A CPU build with a print in
s_ib_flux_wall_prepareshows species 9 (OH) as the limiting species in those cells (consumption ~ 7e-46 againstrhoY~ 2e-46 kg/m^3 per step). Fields at 2e-5 s differ between the two builds by up to 1e-4 relative in the mixture density and 1e-3 in the energy and x-momentum, far above the 1e-6 test tolerance, so the new test fails on master.2. Inert flux wall (commit 2)
Periodic square array of fixed isothermal inert cylinders (area fraction 0.2), carbon-gas mechanism of
examples/2D_ibm_reacting_surfacewith reactions off,thermal_bc = 1, no-slip, body-force driven (bf_x), Re_m = 100 and Mach 0.3 at the start (u_mean 110-190 m/s as the gas heats; the Mach number drifts between about 0.2 and 0.4), 10 nominal flow-throughs (L/U0), WENO5 + HLLC + RK3, A100, 16/24/32/48 cells per diameter. Drag is the x-force on the body and Q the heat rate into the body (ib_force_wrt,ib_heat_wrt), both averaged over the last 5 flow-throughs (the flow is unsteady: the drag standard deviation is 3-4% of its mean, Q about 15%). Errors are relative to the 48 cells/D run with the option; master at 48 cells/D is within 0.3% (drag) and 1.8% (Q) of it, so that reference is not an artifact of the option.Cold wall (gas 600 K, wall 300 K):
The drift is the mean over the 10 flow-throughs; on master at 16 cells/D it is +2.1% in the first and +0.9% in the last. With the option the gas mass is constant to round-off. Absolute values at 48 cells/D with the option: drag 0.1821 N/m, Q 54.98 W/m.
Neutral wall (gas 300 K, wall 300 K, heated only by dissipation):
3D (periodic cubic array of spheres, volume fraction 0.2, cold wall, 8 flow-throughs, A100):
(Two resolutions only, so no convergence statement for 3D; WENO5 needs 26 or more cells per direction, so 16 cells/D does not run.)
What this shows, and what it does not:
s_ib_flux_wall_prepare) uses no per-thread arrays, (removing the arrays is what changed it; I assume the amdflang OpenMP-offload build put them in scratch memory, I did not inspect the generated code): on the MI300A at 96 cells/D it takes 0.030 ms per call (9.0 ms of 100 steps, 1.0% of kernel time) against 3.69 ms per call (1107 ms, 57%) with the arrays, where the option cost +80% per step (9.6 -> 17.4 ns/gp/eq/rhs, two repeats). With the kernel as committed the option is within run-to-run noise on the MI300A (10.66 and 10.70 on; 9.66 and 11.59 off ns/gp/eq/rhs). On the A100 (nvfortran 25.11, GPU 0 otherwise idle, 96 cells/D, 100 fixed steps, three repeats each, interleaved) it is +10.5%: 5.30 (5.29-5.31) off, 5.85 (5.81-5.89) on ns/gp/eq/rhs. The earlier A100 sweep at 16-48 cells/D, with other jobs on the same GPUs (indicative only), gave 173/141, 72.8/60.9, 38.3/36.9, 18.5/16.1 ns/gp/eq/rhs (branch/master), i.e. +4% to +23%.Tests
New (all
Chemistrylabel,override_tol = 1e-6for the cross-vendor single-precision output, goldens generated on the MI300A):2D -> Chemistry -> IBM Reacting Surface -> Flux Wall -> Trace Species(3F5FE078): the reacting cylinder on 48 x 48.2D -> Chemistry -> IBM Inert Surface -> Flux Wall(8B2A7F3E): the same cylinder withsurface_reaction = 0in the freestream.2D -> Chemistry -> IBM Inert Surface -> Flux Wall -> Closed Box(FAF04D7A): inert cylinder in the periodic box of2D_ibm_wall_mass_budget.test_case_validator.py: inert no-slip bodies accepted; slip and blowing inert bodies alone rejected.Results at the head of this branch (
./mfc.sh test --only ..., one rank):--gpu mp)--gpu acc)ChemistryIBMChemistryandIBMtogetherFlux Wall(C1F0BCB1), final tolerancesNo existing golden was regenerated or changed on either vendor.
./mfc.sh formatleaves everything unchanged and./mfc.sh precheckpasses.Limitations
--single,--mixed. The A100 binaries were built before a comment and whitespace reformat ofm_ibm.fppandm_chemistry.fpp(the code is the same; I did not rebuild). Commit 1 on its own was not built or tested separately; itsm_ibm.fppis the theta hunk of the final file.igr, cylindrical coordinates andib_cutvol(rejected as before). Only one mechanism (the carbon/air one) was used; 3D has one sphere configuration;ib_wall_order = 2andib_ip_interp = 2were not combined with the new wall.thetathere.2D_ibm_wall_mass_budgetREADME table are updated to the numbers measured here (defect). Its N2-leak column is not updated: I measure -4e-9 to -5e-9 of the gasified mass for it on both master and branch, against the 1e-10 listed, and did not find why.This PR was made with Claude Code (AI).