Re: [Gfs-users] Gerris parallel performance

[Rob H. <rob...@gm...>]

Stephane,

Sorry for the belated response. I am currently trying to profile the
new cached box traversal patch in gerris-devel. Unfortunately, this
version of the code is causing crashes for me during the initial
domain partitioning. I am using this very simple case (with -s 5, -p
4):

1 0 GfsSimulation GfsBox GfsGEdge {} {
  GfsTime { iend = 0 }
  GfsSolid (difference(x*x + y*y - .1*.1,z-.45))
  GfsRefineSolid 8
  GfsOutputTime { istep = 1 } stdout
  GfsOutputSimulation { istep = 1 } Output-%06ld.gfs { binary = 0 }
  GfsInit {} { U = 1 }
}
GfsBox { left = GfsBoundaryInflowConstant 1 right = GfsBoundaryOutflow }


And I get the following messages that repeat many times:

(process:9053): Gts-CRITICAL **: gts_eheap_remove: assertion `p ==
pdata[i - 1]' failed
(process:9053): Gts-CRITICAL **: gts_eheap_remove: assertion `i > 0 &&
i <= heap->elts->len' failed
....
Gts:ERROR:pgraph.c:29:gnode_split_destroy: assertion failed:
(GTS_SLIST_CONTAINEE (ns->n)->containers == NULL)
[rehlin64:09053] *** Process received signal ***
[rehlin64:09053] Signal: Aborted (6)
[rehlin64:09053] Signal code:  (-6)
[rehlin64:09053] [ 0] /lib/libpthread.so.0(+0xf8f0) [0x7f99272b88f0]
[rehlin64:09053] [ 1] /lib/libc.so.6(gsignal+0x35) [0x7f9926f59a75]
[rehlin64:09053] [ 2] /lib/libc.so.6(abort+0x180) [0x7f9926f5d5c0]
[rehlin64:09053] [ 3] /lib/libglib-2.0.so.0(g_assertion_message+0x144)
[0x7f992853ebb4]
[rehlin64:09053] [ 4] /lib/libglib-2.0.so.0(+0x65130) [0x7f992853f130]
[rehlin64:09053] [ 5] /usr/local/lib/libgts-0.7.so.5(+0x4a246) [0x7f9928c0e246]
[rehlin64:09053] [ 6] /usr/local/lib/libgts-0.7.so.5(+0x492d1) [0x7f9928c0d2d1]
[rehlin64:09053] [ 7]
/usr/local/lib/libgts-0.7.so.5(gts_graph_bisection_new+0xd2)
[0x7f9928c0f9b2]
[rehlin64:09053] [ 8]
/usr/local/lib/libgts-0.7.so.5(gts_graph_recursive_bisection+0xa4)
[0x7f9928c0fb54]
[rehlin64:09053] [ 9] gerris3D(main+0xd1e) [0x402b8e]
[rehlin64:09053] [10] /lib/libc.so.6(__libc_start_main+0xfd) [0x7f9926f44c4d]
[rehlin64:09053] [11] gerris3D() [0x401d99]
[rehlin64:09053] *** End of error message ***


Can you take a quick look and see of something obvious might be
causing these messages. I am very interested in the parallel
performance resulting from the cached box traversal, and I will
continue my testing as soon as I can get these partitioning errors
resolved.

Also, there is still a bug that prevents use of "-p 1" that I
described in a previous post.

Any help on this is greatly appreciated,

-Rob


On Mon, Mar 28, 2011 at 3:20 PM, Stephane Popinet <s.p...@ni...> wrote:
> Hi Rob,
>
>> This is good news. I don't suppose that would be as simple as just
>> removing the qsort(...) call from domain_foreach? Perhaps a flag could
>> be introduced to switch between a deterministic or non-deterministic
>> run? I would be very interested to repeat these profiling studies
>> using unsorted box traversals.
>
> This should work but I think the attached patch:
>
> Mon Mar 28 22:10:05 CEST 2011  Stephane Popinet <s.p...@ni...>
>  * Optimised domain box traversal by caching sorted array.
>
> does better. You can test it by getting (or pulling all patches from)
> the latest gerris-devel branch:
>
> % darcs get http://patch-tag.com/r/popinet/gerris-devel
> % cd gerris-devel
> % darcs apply traversal-patch
> % sh autogen.sh --prefix=...
> % make && make install
>
> Please tell me if this fixes some of your problems (this won't fix the
> slow multigrid convergence though...)
>
> cheers
>
> Stephane
>
> ------------------------------------------------------------------------------
> Create and publish websites with WebMatrix
> Use the most popular FREE web apps or write code yourself;
> WebMatrix provides all the features you need to develop and publish
> your website. http://p.sf.net/sfu/ms-webmatrix-sf
>
> _______________________________________________
> Gfs-users mailing list
> Gfs...@li...
> https://lists.sourceforge.net/lists/listinfo/gfs-users
>
>