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

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

Hello,

Mirko Myllykoski, le lun. 01 juin 2020 13:55:36 +0200, a ecrit:
> On 2020-05-20 19:04, Samuel Thibault wrote:
> > No, the shape seems relatively nice enough. But perhaps put printfs in
> > the loop of _starpu_regression_non_linear_power to check that it does
> > get called and whether it perhaps diverges to nan.
> 
> I investigated the _starpu_regression_non_linear_power function. It appears
> that some list entities (1st argument) are invalid in the sense that the
> "mean" field is 0.0. This causes a situation where "cmax" becomes zero and
> NaNs follow...

Ok. When that happens, what are the values of the other fields?  We have
some code that tries to get rid of outliers, which just flushes the
content of the entry, in which case the mean will indeed be reset to
0.0, but also all other values (sum, sum2, nsample).

> I modified the find_list_size and dump_list functions so that the invalid
> list entries are ignored. This fixes the fitting problem. However, I am
> somewhat concerned over the fact that some of the ignored list entities
> correspond tasks that are not that small.

Ignoring the entries that have nsample == 0 is definitely to be done, I
have now added that.

> Could there be something wrong with the time measurement?

That's very little probable compared to the flush case.

Thanks!
Samuel