Laplace performance test #1908
Laplace performance test #1908
Conversation
|
Looks good to me. Will have to be careful about cases where we might want to provide an initial guess (iterative solvers) as in real cases we would potentially have a decent (or not) guess, whilst here we'd always have the same guess (or a perfect guess if we reused the previous result). |
|
|
||
| MZ = 32 # Z size | ||
|
|
||
| grid = "../../../tests/integrated/test-laplace/test_laplace.grd.nc" |
There was a problem hiding this comment.
I think the information in these grid files is pretty trivial. Maybe better to put the variables into [mesh] section here? Might even just be defaults dx=dy=dz=g11=g22=g33=1, g12=g13=g23=0 at a guess.
There was a problem hiding this comment.
I checked the gridfile, and really all that's in it that's non-default is nx and ny, so can replace with
nx = 12
ny = 16
in the [mesh] section. Actually for a performance test, probably want to increase those numbers a bit... I guess ny does not really matter, effectively it just multiplies the number of iterations and averages over a few slightly different inputs, but a nx and mz in the range 100-200 is probably more representative.
There was a problem hiding this comment.
I think this is following the style of the other performance tests/examples where a scan over grid size is one of the use cases (to look at weak scaling as well as strong scaling etc.). In the other performance cases I think this is included as a part of the script used to run the scan.
There was a problem hiding this comment.
Thanks both. I was following the iterator/make_scaling_example.sh style in runtest, but I stripped out the grid size scan. I can put that back in if people prefer?
(Also, I went with small numbers just to have something that runs in a few seconds on O(1) cores by default.)
There was a problem hiding this comment.
If you wanted to be fancy you could add some arguments to the script so you can ask for a grid size scan, proc scan or both.
|
Should we also backport this? Might be handy to see how the performance has changed between 4.4 and 5.0 |
Yep, will do. |
d7919
added
the
backport candidiate
Adds a simple performance benchmark for timing Laplacian inversion.