Re: [Starpu-devel] About the prediction model of data transfer time

[Samuel Thibault <samuel.thibault@ens-lyon.org>]

Lennox Wu, le Wed 02 Jul 2014 16:48:52 +0800, a écrit :
> I didn't find the prediction model from your papers(Was it written 
> somewhere?),

The paper about performance models is http://hal.inria.fr/inria-00421333

> I  found the function  "starpu_transfer_predict", is it the prediction 
> model?

This is the data transfer prediction model.

> I can't understand the meaning behind the fomular
> "2*(topology->ncudagpus+topology->nopenclgpus)",
> could anybody explain it?

The data transfer prediction is very rough: since the measurement is
done off-line, we divide by the number of the GPUs to account for
contention in a quite rough manner.

The task prediction models are in src/core/perfmodel,
starting from starpu_model_expected_perf, which
ends up for instance in perfmodel_history.c's
_starpu_history_based_job_expected_perf for the simple average over
history, or _starpu_regression_based_job_expected_perf for the
regression-based model, etc.

Samuel