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

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

Mirko Myllykoski, le mar. 09 juin 2020 12:44:00 +0200, a ecrit:
> In particular, you can see that the "c" parameter is quite small for the
> CUDA implementation:

Indeed, and also for the CPU case c is bogus, even if to a lesser
extent. Looking at _starpu_regression_non_linear_power, I see that it
starts its dichotomy for c in [0, min(averages)]. But with the noisy
results that you have, taking the minimum is not a good idea: e.g. in
the CUDA case the minimum average is around 0.004-0.005, but only a few
of the averages have this value. Taking the minimum was meant to make
the dichotomy faster, but with noisy data it doesn't seem like a good
idea. We still need some initial maximum value for the dichotomy, but we
could use the maximum average, or even a coefficiented average of the
average.

Could you give a try to converting find_list_min into find_list_max, and
using that for the initial value of cmax?

Samuel