Accéder au contenu.
Menu Sympa

starpu-devel - Re: [Starpu-devel] Performance profiling of tile low rank cholesky (HiCMA) with StarPU on distributed memory systems

Objet : Developers list for StarPU

Archives de la liste

Re: [Starpu-devel] Performance profiling of tile low rank cholesky (HiCMA) with StarPU on distributed memory systems


Chronologique Discussions 
  • From: Kadir Akbudak <kadir.akbudak@kaust.edu.sa>
  • To: Hatem Ltaief <hatem.ltaief@kaust.edu.sa>
  • Cc: "starpu-devel@lists.gforge.inria.fr" <starpu-devel@lists.gforge.inria.fr>, Aleksandr Mikhalev <aleksandr.mikhalev@kaust.edu.sa>, "Sameh M. Abdulah" <sameh.abdulah@kaust.edu.sa>, Aniello Esposito <esposito@cray.com>
  • Subject: Re: [Starpu-devel] Performance profiling of tile low rank cholesky (HiCMA) with StarPU on distributed memory systems
  • Date: Sun, 13 May 2018 11:15:32 +0300
  • Authentication-results: mail2-smtp-roc.national.inria.fr; spf=None smtp.pra=kadir.akbudak@kaust.edu.sa; spf=Pass smtp.mailfrom=kadir.akbudak@kaust.edu.sa; spf=None smtp.helo=postmaster@mail-qk0-f182.google.com
  • Ironport-phdr: 9a23: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
  • List-archive: <http://lists.gforge.inria.fr/pipermail/starpu-devel/>
  • List-id: "Developers list. For discussion of new features, code changes, etc." <starpu-devel.lists.gforge.inria.fr>

Dear All,

I have attached an experimental case:

* Parallel cholesky algorithm
* 1080K-by-1080K matrix, best block size out of 4 block sizes is reported
* different number of nodes (16, 32, 64,128,256, 512 and 1024; one MPI
rank per node)
* different number of threads per NODE (1,2,4,8,16,31; number of
physical cores is 32)
* different schedulers: lws, prio and ws

My question is:

The brown line (time for 512 nodes) should not go up while the number
of threads per node increases. How can I achieve this?

Thanks,
Kadir


On Wed, May 2, 2018 at 9:44 PM, Hatem Ltaief <hatem.ltaief@kaust.edu.sa>
wrote:
> Kadir should update soon on switching to ws or lws, in situations where
> cores don’t have enough work (i.e., in strong scaling mode of operations).
> Thx,
> H
>> On May 2, 2018, at 8:30 PM, Samuel Thibault <samuel.thibault@inria.fr>
>> wrote:
>>
>> Aniello Esposito, le mer. 02 mai 2018 17:08:48 +0000, a ecrit:
>>> As Hatem already emphasized, the main point here is the behavior of
>>> threads
>>> in the presence of not enough work. We would have expected no real
>>> difference if even
>>> too many threads are used.
>>
>> Sure, but that's precisely where a centralized scheduler like prio would
>> suffer from all the idle threads trying to pick up work from the central
>> place, tasks getting executed randomly on various CPUs, etc.
>>
>> Samuel
>

--

This message and its contents, including attachments are intended solely
for the original recipient. If you are not the intended recipient or have
received this message in error, please notify me immediately and delete
this message from your computer system. Any unauthorized use or
distribution is prohibited. Please consider the environment before printing
this email.

Attachment: lws.pdf
Description: Adobe PDF document

Attachment: prio.pdf
Description: Adobe PDF document

Attachment: ws.pdf
Description: Adobe PDF document




Archives gérées par MHonArc 2.6.19+.

Haut de le page