From: John P. <pet...@cf...> - 2004-03-19 21:24:44
|
KIRK, BENJAMIN (JSC-EG) (NASA) writes: > > Before I go off and potentially break everything I'd like to have some > reliable way to get back to where we are now. So, we need some regression > tests. Since I'm trying to balance finishing my degree with full-time work > and this parallel implementation I'd like to solicit help from the team. My favorite regression tests have always been recovering the (hopefully correct!) asymptotic error behavior for uniformly refined grids. Method of manufactured solutions (for linear problems) to generate exact solutions or special well-known cases (Kovasznay flow) for non-linear problems would be needed to compute the error. I have wanted to set up a problem involving Kovasznay flow for some time, so now hopefully I will get around to it. We can also use object-orientation to (potentially) make this easier for us. For example, if we create a RegressionTest object which takes an EquationSystems, then the object could solve the system on a sequence of grids, using different IO formats, with and without AMR, with different solvers, single and multi-processor, etc. etc. That way the test code could be written once and applied to several different systems. I suggest we *not* take the same approach as in the examples where we numbered them ex1, ex2, ... etc. and give each test a unique name instead. The Makefiles may be more complex, but that way you don't have to worry if somebody is going to check in test7 before you get around to it. I suggest Daniel, Steffen, et. al. focus on regression tests for the complex parts of the code. Personally, I have not configured complex variables "on" for a long time, and could not vouch for their correctness ;) -John |