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

[Mirko Myllykoski <mirkom@cs.umu.se>]

Hi,

On 2020-06-03 20:41, Samuel Thibault wrote:

> 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).

How/where is this outlier filtering performed? I am asking because I 
have been getting some weird results with a (consumer-level) GPU. Please 
see the attachment.

In particular, you can see that the "c" parameter is quite small for the 
CUDA implementation:
        Non-Linear: y = a size ^b + c
                a = 3.220354e-06
                b = 7.817886e-01
                c = 1.084335e-03
        Non-Linear: y = a size ^b + c
                a = 1.016593e-03
                b = 3.854069e-01
                c = 2.572536e-10

That is, the model is very close to being linear and clearly does not 
fit to the data.

I believe this happens because the fitting algorithm fails to find the 
correct "c" parameter. Based on the data, the "c" parameter should 
probably be ~0.06. However, since the "c" parameter is always smaller 
than "cmax" and "cmax" is ~0.006, the "c" parameter is always going to 
be far outside the range where the fitting algorithm can actually 
converge to something useful.

 - Mirko

By sending an email to Umeå University, the University processes your 
personal data. For more information, please read: https://www.umu.se/en/gdpr

Attachment: starpu_starneig_right_gemm_update_pm.banana.eps
Description: image/eps

Attachment: starpu_starneig_right_gemm_update_pm.banana.gp
Description: application/gnuplot