[aspect-devel] performance

[Thieulot, C. (Cedric)]

Hi all,

back at work on Aspect and performance related issues. Following your last email's advice to
use a different trilinos, I have installed trilinos-11.0.3 (we used 10.4.2 so far) yesterday and I am happy to report that it makes a difference:

setup reinit                                has gone from           3.55s   to   1.53s
setup system matrix                 has gone from      214s        to    68.8s
setup system preconditioner  has gone from         95.2s    to    21.2s

Q1: are these numbers normal in your opinion ? still too high ?

However, it has also had a nasty consequence on the Stokes timings:

Assemble Stokes system           has gone from 314s to 445s
Build Stokes preconditioner     has gone from 138s to 250s
Solve Stokes system                  has gone from 81.4 to 115s

(Assemble composition system has remained the same at about 520s )
(Assemble temperature system has remained the same at about 520s )

I have had a look at the assembly routine and put there a number of "computing_timer" commands.
It turns out that the part between line 955 and line 998 (the WorkStream:: run … part)  takes 260s
and that the part after it "system_matrix.compress();" takes 363s.

Q2: would you have any idea of what is happening and how we could solve this in the briefest delays ?
(EGU spring meeting is in less than three weeks)

thanks,

Cedric.

ps: I hereby attach again a typical output, which is obtained on threads.




[cid:0136F42E-C0EB-46E9-B601-E3C818D5CE87 at geo.uu.nl]
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://geodynamics.org/pipermail/aspect-devel/attachments/20130320/38b6539a/attachment-0001.htm 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: opla_vkk3D_trilinos11.png
Type: image/png
Size: 175539 bytes
Desc: opla_vkk3D_trilinos11.png
Url : http://geodynamics.org/pipermail/aspect-devel/attachments/20130320/38b6539a/attachment-0001.png