Scotch 6.0 is available.

[QuLogic]

I don't use it, but we should check whether our code can build against it and maybe update the bundled copy.

Here are some release notes:

Changes from version 5.1

• Direct k-way graph partitioning and static mapping methods are now available. They are less expensive than the classical dual recursive bipartitioning scheme, and improve quality on average for numbers of parts above a few hundreds. Another new method aims at reducing load imbalance in the case of source graphs with highly irregular vertex weights; see Section 3.3, page 13. Users willing to keep using the old recursive bipartitioning strategies of the 5.x branch can create default strategies with the SCOTCH STRATRECURSIVE flag set, in addition to other flags; see Section 8.3.1, page 59 for further information.
• Graph repartitioning and static re-mapping features are now available; see Sections 8.7.2 to 8.7.7, starting from page 93.
• The clustering capabilities of Scotch can be used more easily from the command line and library calls; see Section 7.3.6 and Section 8.15.2.
• A new set of routines has been created in order to compute vertex-separated, k-way partitions, that balance the loads of the parts and of the separator vertices that surround them; see Sections 8.3.4 and 8.7.5.
• A new labeled tree-leaf architecture has been created, for nodes that label cores in non increasing order. See Section 6.4.2, page 26 for the description of the ltleaf target architecture.
• Memory footprint measurement routines are now available to users; see Section 8.19, page 144.
• Key algorithms are now multi-threaded. See the installation file INSTALL.txt in the main directory for instructions on how to compile Scotch with thread support enabled.

[komatits]

probably not urgent to switch to 6.0, since 5.2 works great.

Dimitri.

On 11/14/2013 10:59 PM, Elliott Sales de Andrade wrote:

I don't use it, but we should check whether our code can build against it and maybe update the bundled copy.

Here are some release notes:

  Changes from version 5.1

• Direct k-way graph partitioning and static mapping methods are now available. They are less expensive than the classical dual recursive bipartitioning scheme, and improve quality on average for numbers of parts above a few hundreds. Another new method aims at reducing load imbalance in the case of source graphs with highly irregular vertex weights; see Section 3.3, page 13. Users willing to keep using the old recursive bipartitioning strategies of the 5.x branch can create default strategies with the SCOTCH STRATRECURSIVE flag set, in addition to other flags; see Section 8.3.1, page 59 for further information.
• Graph repartitioning and static re-mapping features are now available; see Sections 8.7.2 to 8.7.7, starting from page 93.
• The clustering capabilities of Scotch can be used more easily from the command line and library calls; see Section 7.3.6 and Section 8.15.2.
• A new set of routines has been created in order to compute vertex-separated, k-way partitions, that balance the loads of the parts and of the separator vertices that surround them; see Sections 8.3.4 and 8.7.5.
• A new labeled tree-leaf architecture has been created, for nodes that label cores in non increasing order. See Section 6.4.2, page 26 for the description of the ltleaf target architecture.
• Memory footprint measurement routines are now available to users; see Section 8.19, page 144.
• Key algorithms are now multi-threaded. See the installation file INSTALL.txt in the main directory for instructions on how to compile Scotch with thread support enabled.

— Reply to this email directly or view it on GitHub https://github.com/geodynamics/specfem3d/issues/71.

Dimitri Komatitsch CNRS Research Director (DR CNRS), Laboratory of Mechanics and Acoustics, UPR 7051, Marseille, France http://komatitsch.free.fr

[QuLogic]

Sure, this is just a reminder "for the future".

[mpbl]

see also issue #19

[mpbl]

@QuLogic Great idea. As discussed with @komatits, @danielpeter and @luet, you may want to remove scotch from specfem3d repository and advise people to download them by themselves. The autotools scripts might need to be changed accordingly.

[komatits]

Hi all,

This is not what I meant, I meant leaving the Scotch source code in a SPECFEM3D directory but removing it from buildbot, so that warnings about portability issues or unused variables in that library are ignored in the buildbot test.

I think it is much better to keep providing the full SCOTCH source code with our package, this way users do not need to download anything else. We can, since SCOTCH is open source.

Thanks, Dimitri.

On 23/01/2014 19:56, Matthieu Lefebvre wrote:

@QuLogic https://github.com/QuLogic Great idea. As discussed with @komatits https://github.com/komatits, @danielpeter https://github.com/danielpeter and @luet https://github.com/luet, you may want to remove scotch from specfem3d repository and advise people to download them by themselves. The autotools scripts might need to be changed accordingly.

— Reply to this email directly or view it on GitHub https://github.com/geodynamics/specfem3d/issues/71#issuecomment-33155817.

Dimitri Komatitsch CNRS Research Director (DR CNRS), Laboratory of Mechanics and Acoustics, UPR 7051, Marseille, France http://komatitsch.free.fr

[mpbl]

see commit b683d0a2f9d7bdc237d40a0a1166b8bb01a67a98

[komatits]

It seems that switching to scotch v6.0 was done in the 3D code only; should be done in SPECFEM2D as well, for consistency (SPECFEM2D still uses scotch 5.2)

[mpbl]

@komatits Can you do it -- I never used the 2D code? Thanks.

[komatits]

Yes, I could, but I am not sure scotch_6.0.1rc2 is reliable because I get the following warning from Intel icc:

icc -O3 -DCOMMON_FILE_COMPRESS_GZ -DCOMMON_PTHREAD -DCOMMON_RANDOM_FIXED_SEED -DSCOTCH_RENAME -DSCOTCH_RENAME_PARSER -DSCOTCH_PTHREAD -Drestrict=__restrict -pthread -c library_graph_map_view.c -o library_graph_map_view.o

library_graph_map_view.c(154): warning #592: variable "grafptr" is used before its value is set vmlotax = (Gnum *) vmlotab - grafptr->baseval;

Thus let me keep using v5.2 in the 2D code for now.

(I will also email F Pellegrini about this bug)

On 06/03/2014 17:34, Matthieu Lefebvre wrote:

@komatits https://github.com/komatits Can you do it -- I never use the 2D code? Thanks.

— Reply to this email directly or view it on GitHub https://github.com/geodynamics/specfem3d/issues/71#issuecomment-36906726.Web Bug from https://github.com/notifications/beacon/5817129__eyJzY29wZSI6Ik5ld3NpZXM6QmVhY29uIiwiZXhwaXJlcyI6MTcwOTc0Mjg5MiwiZGF0YSI6eyJpZCI6MjAyOTk5NDN9fQ==--11c1e02503d304ca580d79a6af26e105a0d9db99.gif

Dimitri Komatitsch CNRS Research Director (DR CNRS), Laboratory of Mechanics and Acoustics, UPR 7051, Marseille, France http://komatitsch.free.fr

[komatits]

Done by Matthieu.

[QuLogic]

Done by Matthieu.

I don't see anything done in the 2D code?

[komatits]

True: done by Matthieu in the 3D code, purposely not done in the 2D code because we found some warnings and a potential bug in Scotch6_rc1 while trying to compile it with SPECFEM2D.

(I would thus suggest switching back to Scotch5 in the 3D code; it is probably not good to use *_rc versions of libraries, we should use stable releases only)

On 18/03/2014 01:34, Elliott Sales de Andrade wrote:

Done by Matthieu.

I don't see anything done in the 2D code?

— Reply to this email directly or view it on GitHub https://github.com/geodynamics/specfem3d/issues/71#issuecomment-37888106.

Dimitri Komatitsch CNRS Research Director (DR CNRS), Laboratory of Mechanics and Acoustics, UPR 7051, Marseille, France http://komatitsch.free.fr