Accéder au contenu.
Menu Sympa

starpu-devel - [Starpu-devel] memory overhead

Objet : Developers list for StarPU

Archives de la liste

[Starpu-devel] memory overhead


Chronologique Discussions 
  • From: Xavier Lacoste <xavier.lacoste@inria.fr>
  • To: starpu-devel@lists.gforge.inria.fr
  • Cc: Mathieu Faverge <Mathieu.Faverge@inria.fr>, Pierre Ramet <ramet@labri.fr>
  • Subject: [Starpu-devel] memory overhead
  • Date: Fri, 6 Dec 2013 14:52:59 +0100
  • 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>

Hello,

I'm studying memory overhead with StarPU/ParSEC/native scheduler in PaStiX and i noticed a slightly bigger oerhead with StarPU.
I also noticed a very very large number of allocation with StarPU 1.1 (r11396M) (from eztrace report).

Here are two figures, the first one is the memory peak with the schedulers overhead in blue/red/beige, and the ratio compared to PaStiX peak on top, the second the number of allocations obtained during the run.

http://img833.imageshack.us/img833/816/k1sq.jpg

http://img843.imageshack.us/img843/7166/fa3x.jpg

This is obtained with the direct linear solver PaStiX on 9 different matrices tests cases, on 12 cores on mirage.

Have you go an idea of what could explain this overhead (+8% compared to PaStiX peak but much more if we don't take into account common data structures, which may be a more interesting metric....) and the huge number of calls to malloc().

Regards,

XL.



----------------------------------------
Xavier Lacoste
INRIA Bordeaux Sud-Ouest
200, avenue de la Vieille Tour
33405 Talence Cedex
Tél : +33 (0)5 24 57 40 69











Archives gérées par MHonArc 2.6.19+.

Haut de le page