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

[Olivier Aumage <olivier.aumage@inria.fr>]

Hi Mawussi,

Sure! :-)

Best regards,
--
Olivier

> Le 25 sept. 2017 à 12:18, Mawussi Zounon <mawussi.zounon@manchester.ac.uk> 
> a écrit :
> 
> Hi Oliver,
> 
> It can worth including the StarPU initialization in at least one routine, 
> for the sake of comparison, 
> just to evaluate the gain of the Samuel's improvement.
> 
> Best regards,
> --Mawussi 
> ________________________________________
> From: Olivier Aumage [olivier.aumage@inria.fr]
> Sent: Monday, September 25, 2017  5:13 PM
> To: Mawussi Zounon
> Cc: Samuel Thibault; Negin Bagherpour; starpu-devel@lists.gforge.inria.fr; 
> Jakub Sistek
> Subject: Re: [Starpu-devel] Performance issue of StarPU on Intel 
> self-hosted KNL
> 
>> Le 25 sept. 2017 à 12:10, Mawussi Zounon <mawussi.zounon@manchester.ac.uk> 
>> a écrit :
>> 
>> Thanks Samuel,
>> 
>> I will perform a quick benchmark then be back to you.
>> 
> Hi Mawussi,
> 
> Even with a single topology load from hwloc now, it is still not advised to 
> perform a starpu_init()/starpu_shutdown() on every PLASMA call. :-)
> 
> Best regards,
> --
> Olivier
> 
>> Best regards,
>> --Mawussi
>> ________________________________________
>> From: Samuel Thibault [samuel.thibault@inria.fr]
>> Sent: Monday, September 25, 2017 1:16 PM
>> To: Olivier Aumage; Mawussi Zounon; Negin Bagherpour; 
>> starpu-devel@lists.gforge.inria.fr; Jakub Sistek
>> Subject: Re: [Starpu-devel] Performance issue of StarPU on Intel 
>> self-hosted KNL
>> 
>> Samuel Thibault, on ven. 22 sept. 2017 11:48:12 +0200, wrote:
>>> 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.
>> 
>> I have fixed this.
>> 
>> Samuel
>