[aspect-devel] Aspect Timer

Timo Heister heister at math.tamu.edu
Thu May 3 16:20:25 PDT 2012


Hi Eric,

Interesting. I checked in change to record the timings of creating the
snapshot (the output/mesh file).
Can you check it that takes all the time? How big is the file for you?

On Thu, May 3, 2012 at 5:00 PM, Eric Heien <emheien at ucdavis.edu> wrote:
> I just ran a test of Aspect on Lonestar using 48 processors with the box.prm
> sample (using more mesh refinement) and I got the following timings:
>
> +---------------------------------------------+------------+------------+
> | Total wallclock time elapsed since start    |       429s |            |
> |                                             |            |            |
> | Section                         | no. calls |  wall time | % of total |
> +---------------------------------+-----------+------------+------------+
> | Assemble Stokes system          |       131 |      1.25s |      0.29% |
> | Assemble temperature system     |       131 |      3.06s |      0.71% |
> | Build Stokes preconditioner     |        30 |      1.71s |       0.4% |
> | Build temperature preconditioner|       131 |     0.181s |     0.042% |
> | Solve Stokes system             |       131 |       4.6s |       1.1% |
> | Solve temperature system        |       131 |      1.68s |      0.39% |
> | Postprocessing                  |       129 |      17.3s |         4% |
> | Refine mesh structure, part 1   |        27 |     0.913s |      0.21% |
> | Refine mesh structure, part 2   |        27 |      0.16s |     0.037% |
> | Setup dof systems               |        28 |      1.22s |      0.28% |
> +---------------------------------+-----------+------------+------------+
>
> In case the formatting is messed up, basically it seems the total wall time
> (429s) is about 13x greater than the sum of the sections.  In other words,
> 397 seconds of run time are unaccounted for out of 429 total seconds.  Are
> there any sections that could potentially account for this that are not yet
> timed?
>
> I know the tracers take a long time so I turned them off but I still see
> this discrepancy (plus they should be accounted for under "Postprocessing").
>  One possible hint is that it seems the "mesh" file is being altered
> extensively during checkpoints - perhaps this is related?
>
> Thanks for any ideas,
>
> -Eric
>
>
> _______________________________________________
> Aspect-devel mailing list
> Aspect-devel at geodynamics.org
> http://geodynamics.org/cgi-bin/mailman/listinfo/aspect-devel
>



-- 
Timo Heister
http://www.math.tamu.edu/~heister/


More information about the Aspect-devel mailing list