Re: [Starpu-devel] Performance issue of StarPU on Intel self-hosted KNL

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

Hello,

Olivier Aumage, on ven. 22 sept. 2017 09:12:27 +0200, wrote:
> The main reason for the initialization cost on KNL is the large number of 
> cores and the fact that those cores are individually much less powerful and 
> advanced than Broadwell's cores. Initializing libhwloc on the KNL takes 
> about 0.5 to 1 second.

And we do it twice, once for the whole architecture, and a second
time in _starpu_build_tree to get only the structure of the
architecture. Perhaps that second call could be avoided, it's probably
less expensive to make build_tree cope with structure-useless hwloc
object, than asking hwloc discovery a second time.

> Then, the initialization of StarPU itself involves thread creation, 
> synchronization, loading bus sampling data for every core, ...

I don't think it's that much expensive compared to the hwloc
discovery. When there is no GPU, there is no bus sampling, and we
share the performance models for cores. The thread creation and
synchronization is parallelized as much as possible.

Samuel