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

[Mawussi Zounon <mawussi.zounon@manchester.ac.uk>]

Dear Samuel,

We are using StarPU 1.2.2.
In terms of scheduling, I tried eager, and lws, but didn't notice a 
significant difference though lws seems better.
Currently we are  not using priority, but we will need it very soon with LU.
I will try the trunk and be back to you.

Best regards
--Mawussi

  
________________________________________
From: Samuel Thibault [samuel.thibault@inria.fr]
Sent: Monday, September 18, 2017 5:27 PM
To: Mawussi Zounon
Cc: starpu-devel@lists.gforge.inria.fr; Olivier Aumage; Jakub Sistek; Negin 
Bagherpour
Subject: Re: Performance issue of StarPU on Intel self-hosted KNL

Hello,

Which version of StarPU is this?

Which scheduler did you use? The default scheduler in StarPU 1.2 is
still "eager", which is a real bottleneck for 60-y-core systems :)

Using "lws" would probably be a better go.

Trying the StarPU trunk (it is very stable ATM, we will probably branch
to 1.3 sooner or later) could be useful too, its default scheduler is
now lws too.

Also, you might want to try the trunk's --enable-numa option. It is
still very experimental and was not really benchmarked, but possibly it
could help.

Are there priorities defined in your algorithm?

Samuel