Re: [Starpu-devel] Problem with performance model calibration

[Samuel Thibault <samuel.thibault@inria.fr>]

Mirko Myllykoski, le mer. 20 mai 2020 17:13:50 +0200, a ecrit:
> More data but quite noisy. Weirdly, it appears that the code still
> default to the linear models...

More precisely, that it didn't compute the non-linear model. (the linear
one is always computed since it's easy).  Perhaps check how that happens
for tests/perfmodels/non_linear_regression_based.c vs how that doesn't
seem to happen with your application. Normally that happens during
starpu_shutdown in _starpu_deinitialize_registered_performance_models,
in dump_reg_model that calls _starpu_regression_non_linear_power.

> I looks like both CPU and GPU measurement are mixed in the *.out files... Is
> this true?

That might be true, I don't know whether MRM was actually tried with
different archs, possibly some documentation steps or script fixing
needs to be done to separate out the .out files.  At least that would
clearly explain why MRM is lost, since the performances are really
unrelated. At worse you can separate out by hand in two steps with
STARPU_NCPU=0 vs STARPU_NCUDA=0.

> If GPU(s) are detected, the code defaults to dmdas,

Ok, so that's why the graph was indeed showing only CPU dots for small
sizes and only GPU dots for big sizes.

Samuel