Resilience experiments with Rayleigh-Benard convection #495
Conversation
|
Hm, is anything of this tested somewhere? |
|
No, that's too expensive, I'm afraid. I had to choose a large resolution to get low residuals in the collocation problem and now these experiments take about 40s+ on my laptop each. Also, they need the reference solution. Computing that takes multiple hours, and checking it in is not great either. |
|
When doing experiments with work-precision plots, I noticed some issues with the reference solution. These are fixed now, there is now a test that the time that is supposed to be reached is actually reached very closely. The result is also more sensitivity to faults in RBC. |
| PROBLEM_PARAMS = {'Rayleigh': 2e4, 'nx': 256, 'nz': 128} | ||
|
|
||
|
|
||
| class ReachTendExactly(ConvergenceController): |
pySDC/projects/Resilience/RBC.py
Outdated
| ax.legend(frameon=False) | ||
|
|
||
|
|
||
| def test_order(t=14, dt=1e-1, steps=6): |
There was a problem hiding this comment.
Sure you want to call this test_xxx ?
…ime#495) * Added basic infrastructure for Rayleigh-Benard in resilience project * Verified order and added configurations for dt-adaptivity and base strategy * Implemented configurations for faults in RBC * Added plot of RBC solution * Limit fault targets in space index * Fix * Fixed fault insertion time for Hot Rod * Changed colormap for RBC plots in the paper * Fixed bug in reference solution * Changed RBC configurations * A few more changes to RBC config * Implemented some tests * Restored behaviour of work_precision script * Small fixes #!!!!!! WARNING: FLAKEHEAVEN FAILED !!!!!!: #:
This PR includes a configuration for Rayleigh-Benard convection (RBC) on a 256x128 grid with Rayleigh number of$2^4$ . Reference solutions are computed for t=20 and t=21. Individual runs for fault experiments then start at t=20 and go to t=21. There is a convergence controller that makes sure the target time is reached to accuracy of $10^{-9}$ , which is sufficiently accurate to use the reference solution. I verified that I get the order of the collocation problem relative to such a reference solution, even if the residual stays at order 1 during the runs. The latest theory is that this is because the problem is perturbed by the method that enforces the boundary conditions. This perturbation is missing in the residual and I haven't found a good way to represent it there (yet). This matters when the perturbation is not small, i.e. the resolution is insufficient in z-direction. The configuration used here, however, is sufficiently resolved for this not to cause any issues.