libmesh-users — General information for libmesh users

Re: [Libmesh-users] getpot bug report

[Roy S. <roy...@ic...>]

On Wed, 11 Oct 2017, simone via Libmesh-users wrote:

> I apologize if this is not the right place to report this problem,

This is the right place to report it, yes.  We had to fork from the
original GetPot after newer versions from the original software author
stopped using an open source license, so at this point libMesh GetPot
is full of new features (such as the namespace option) and probably
new bugs, none of which will be relevant upstream.

> I'm having trouble using GetPot library (not the one included in libMesh
> installation), paired with Libmesh. I have isolated the problem in the
> attached pack,

I'm afraid our email list typically strips attachments.  I think I
understand the problem from your description alone, but if I'm wrong
then you could send to my email address specifically with the
attachment.

> when running it I have a segfault error at the exit of
> the program. I know there's a conflict beetwen getpot class definitions,
> but I don't understand why this occur even without including the libmesh
> getpot.h header.

Probably because it's a link-time error?  Although all the GetPot
functions are declared inline, that keyword doesn't actually *force* the
compiler to inline things!  Glancing at my gcc 7.2 libMesh build, for
instance, "nm --demangle libmesh_opt.so | grep GetPot" shows me
literally dozens of symbols corresponding to compiled GetPot
functions.

In practice, IIRC this means that the linker will pick one version or
the other of each of those symbols when it seems them multiply-defined
in different libraries.  If a particular function (in your case,
probably the destructor) is chosen from version A but ends up dealing
with a class layout defined by version B, then it's going to try to
look for member variables in the wrong places and the result will be a
disaster.

In theory, even if you get lucky, you shouldn't bet on it.  The C++
standard says that you can have more than one definition of an inline
function in your program only so long as all definitions are
*identical*; when you include the same function multiple times with
different definitions, the result is undefined.

> I'm forced to use GetPot library in my program, I guess

The original author's version of it, you mean?

> one possible solution could be to install libMesh library with the
> variable GETPOT_NAMESPACE defined, but I'd like to avoid it because my
> code is part of a library that can be delivered to external users using
> their standard installation of libMesh, and I can't force them to
> recompile their installed package. Is there another possible solution?

Since the C++ standard says you can't have two different definitions
of the same function in the same program, you've logically only got
four possibilities, and none of them look like great options.

1) Use one definition: use the libMesh GetPot.  I assume from "forced"
that this isn't an option.

2) Use two different functions.  This is what libMesh was trying to
enable via the GETPOT_NAMESPACE option.  In hindsight I now see that
we obviously should have defaulted that to "libMesh" rather than
undefined, in our own usage... but if you have external users then
it's probably too late for you for us to do that now.  If I'm wrong
about that let us know.

Conversely, even if the original GetPot author hasn't added the same
functionality, it wouldn't be hard for you to patch in manually... but
I assume if you're "forced" to use his GetPot then you're probably not
allowed to modify it either?

3) Use two different programs.  The part that needs one GetPot
communicates with the part that needs the other via a pipe or shared
memory or some such.  This is almost certainly too much work.

4) Use something outside the C++ standard.  If you're on a system that
gives you additional standards regarding symbol resolution, then you
might possibly make use of those, whether it's via tricks with link
order, manually limiting symbol exports, linker "version scripts",
etc.  The best introduction I can find quickly is here:

https://holtstrom.com/michael/blog/post/437/Shared-Library-Symbol-Conflicts.html

> Another question, about public method: libMesh::GnuPlotIO::write(fname)
>
> It seems to do nothing but raising an exception,

What exception?  From what line?  These are very important details.

> and there's no way to avoid it. Is this the expected behaviour?

Definitely not in general.

> Is there a way to print a 1d mesh on a gnuplot readable file as the
> method is depicted to do in the class documentation?

I certainly believe so.  Four of our example codes do it, and I do see
gnuplot_script and gnuplot_script_data output files for each of them
in my build directory from "make check" yesterday.  Do those work for
you or throw the same exception?
---
Roy

[Libmesh-users] getpot bug report

[simone <sob...@ya...>]

Dear Libmesh developers,

I apologize if this is not the right place to report this problem, or if
it has been already raised by someone else. I searched previous digests
for a solution, but I did not find anything similar.

I'm having trouble using GetPot library (not the one included in libMesh
installation), paired with Libmesh. I have isolated the problem in the
attached pack, when running it I have a segfault error at the exit of
the program. I know there's a conflict beetwen getpot class definitions,
but I don't understand why this occur even without including the libmesh
getpot.h header. I'm forced to use GetPot library in my program, I guess
one possible solution could be to install libMesh library with the
variable GETPOT_NAMESPACE defined, but I'd like to avoid it because my
code is part of a library that can be delivered to external users using
their standard installation of libMesh, and I can't force them to
recompile their installed package. Is there another possible solution?

Another question, about public method: libMesh::GnuPlotIO::write(fname)

It seems to do nothing but raising an exception, and there's no way to
avoid it. Is this the expected behaviour? Is there a way to print a 1d
mesh on a gnuplot readable file as the method is depicted to do in the
class documentation?

I thank you in advance for your answer.

Simone

Re: [Libmesh-users] Lagrange factors

[David K. <dav...@ak...>]

On Wed, Oct 11, 2017 at 5:02 AM, Tobias Möhle <tob...@un...>
wrote:

> Dear all,
> I am trying to extend my current code to add Lagrange multipliers, leading
> to extra terms that couplet certain DOFs in a fixed range of the grid.
> To do so, I added another variable to my system, leading to the
> system-matrix of block-structure:
>
> / H L^T \
> \ L  0     /
> where H is the matrix corresponding to the actual problem and L contains
> the Lagrange constrains.
> Thus, my system.solution contains in the first part the coefficients of
> the particular solution and in the lower part the Lagrange multipliers
> themselves.
> Now my question: is there some function to extract the Lagrange
> multipliers and the actual solution vector (dof-based) separately? Or how
> is it supposed to be done?
>


I suggest you refer to systems_of_equations_ex5, since that uses a Lagrange
multiplier to impose a constraint. Hopefully that answers your question.

David

[Libmesh-users] Lagrange factors

[Tobias M. <tob...@un...>]

Dear all,
I am trying to extend my current code to add Lagrange multipliers, leading to extra terms that couplet certain DOFs in a fixed range of the grid.
To do so, I added another variable to my system, leading to the system-matrix of block-structure:

/ H L^T \
\ L  0     /
where H is the matrix corresponding to the actual problem and L contains the Lagrange constrains.
Thus, my system.solution contains in the first part the coefficients of the particular solution and in the lower part the Lagrange multipliers themselves.
Now my question: is there some function to extract the Lagrange multipliers and the actual solution vector (dof-based) separately? Or how is it supposed to be done?

Many thanks in advance,
Tobias

Re: [Libmesh-users] Conforming adaptivity

[Giacomo R. de S. <gia...@ep...>]

Hey, sure, here is a link to a copy in gdrive
https://drive.google.com/open?id=0B1zRh3hYRviqQjFoa1ZhV0tLWk0

Giacomo


On 10/03/2017 08:48 PM, John Peterson wrote:
>
>
> On Tue, Oct 3, 2017 at 11:27 AM, Rosilho De Souza Giacomo 
> <gia...@ep... 
> <mailto:gia...@ep...>> wrote:
>
>     Hello Dimitrios, developers,
>     I’ve written such a code some time ago, conforming adaptivity
>     using Delaunay triangulation. Is that that you need? It was
>     working well in my case where I had to solve elliptic and
>     parabolic problems with first order continuous galerkin, P1
>     lagrange basis. I’ve sent a link to a downloadable example in this
>     mailing list. Did you have a look? Developers?
>     https://sourceforge.net/p/libmesh/mailman/message/35758424/
>     <https://sourceforge.net/p/libmesh/mailman/message/35758424/>
>
>
> Oops, looks like I said I was going to try and create a branch at the 
> time but never got around to it.
>
> We can't access dropbox links at work for some reason, so if you can 
> reshare the code on Google Drive or some other service that would also 
> be helpful.
>
> -- 
> John

Re: [Libmesh-users] Compiling with clang

[John P. <jwp...@gm...>]

On Tue, Oct 3, 2017 at 1:19 PM, Salazar De Troya, Miguel <
sal...@ll...> wrote:

> Hello
>
> I want to compile libMesh with mvapich2-2.2-clang-4.0.0, but I am getting
> errors like this in the configure stage:
>
> configure:10326: /usr/tce/packages/mvapich2/mvapich2-2.2-clang-4.0.0/bin/mpicxx
> -c -std=c++14   conftest.cpp >&5
> conftest.cpp:79:25: error: no member named 'is_trivially_copyable' in
> namespace 'std'
> << std::is_trivially_copyable<char>::value // Not supported by GCC 4.6.3
> with -std=c++0x
>
> It seems to me that this feature should be supported by clang 4.0.0, but
> it is not recognized, probably because the flags I am passing. Do I need to
> pass other flags?
>

I wouldn't worry about this too much.

For one thing, this is not a fatal error, it just means that your compiler
doesn't fully support the <type_traits> header.

For another thing, libmesh doesn't actually use the
LIBMESH_HAVE_CXX11_TYPE_TRAITS for anything, so it would only matter if you
are somehow using this #ifdef in your app.

That said, I use clang 3.9.0 and it passes this test, so I'm not sure what
might be different with your clang 4.0.0 install.

Also, it looks like you are somehow passing -std=c++14 to the configure
test (manually setting libmesh_CXXFLAGS?) which I haven't tested myself,
but should not make any difference to this particular test.

-- 
John

[Libmesh-users] Compiling with clang

[Salazar De T. M. <sal...@ll...>]

Hello

I want to compile libMesh with mvapich2-2.2-clang-4.0.0, but I am getting errors like this in the configure stage:

configure:10326: /usr/tce/packages/mvapich2/mvapich2-2.2-clang-4.0.0/bin/mpicxx -c -std=c++14   conftest.cpp >&5
conftest.cpp:79:25: error: no member named 'is_trivially_copyable' in namespace 'std'
<< std::is_trivially_copyable<char>::value // Not supported by GCC 4.6.3 with -std=c++0x

It seems to me that this feature should be supported by clang 4.0.0, but it is not recognized, probably because the flags I am passing. Do I need to pass other flags?

Thanks
Miguel



--

Re: [Libmesh-users] Conforming adaptivity

[John P. <jwp...@gm...>]

On Tue, Oct 3, 2017 at 11:27 AM, Rosilho De Souza Giacomo <
gia...@ep...> wrote:

> Hello Dimitrios, developers,
> I’ve written such a code some time ago, conforming adaptivity using
> Delaunay triangulation. Is that that you need? It was working well in my
> case where I had to solve elliptic and parabolic problems with first order
> continuous galerkin, P1 lagrange basis. I’ve sent a link to a downloadable
> example in this mailing list. Did you have a look? Developers?
> https://sourceforge.net/p/libmesh/mailman/message/35758424/


Oops, looks like I said I was going to try and create a branch at the time
but never got around to it.

We can't access dropbox links at work for some reason, so if you can
reshare the code on Google Drive or some other service that would also be
helpful.

-- 
John

Re: [Libmesh-users] Conforming adaptivity

[Rosilho De S. G. <gia...@ep...>]

Hello Dimitrios, developers,
I’ve written such a code some time ago, conforming adaptivity using Delaunay triangulation. Is that that you need? It was working well in my case where I had to solve elliptic and parabolic problems with first order continuous galerkin, P1 lagrange basis. I’ve sent a link to a downloadable example in this mailing list. Did you have a look? Developers?
https://sourceforge.net/p/libmesh/mailman/message/35758424/

The link isn’t working anymore but if you need I can repost it or, Dimitrios, just come to my office ;)

Regards,
Giacomo Rosilho de Souza

—————————————
EPFL SB MATHICSE ANMC
MA C2 637 (Bâtiment MA)
Station 8
CH-1015 Lausanne

Il giorno 3-ott-2017, alle ore 16:44, Roy Stogner <roy...@ic...<mailto:roy...@ic...>> ha scritto:


On Tue, 3 Oct 2017, dim...@ep...<mailto:dim...@ep...> wrote:

I would like to ask if libmesh can support conforming mesh adaptation.

Technically, "can support" - yes.  One of the consequences of the
"toolkit, not a framework" philosophy is that users can do a lot of
things without specific library APIs to do them, and looping through
elements to split some of them would be a good example.  libMesh will
even do a few things to help with that, like reconstructing neighbor
topology for you if you don't do that on the fly yourself.

But "can provide" - no.  Currently the only refinement types built
into the library are isotropic h refinement with hanging nodes,
isotropic p refinement on hierarchic element types, nodal
redistribution, and mixtures of those.

If you do write conforming mesh adaptation code, though, and it's
generic enough to work on others' problems (even if it's only for
triangles/tets), we'd love to incorporate it.
---
Roy

------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org<http://slashdot.org>! http://sdm.link/slashdot
_______________________________________________
Libmesh-users mailing list
Lib...@li...<mailto:Lib...@li...>
https://lists.sourceforge.net/lists/listinfo/libmesh-users

Re: [Libmesh-users] Conforming adaptivity

[Roy S. <roy...@ic...>]

On Tue, 3 Oct 2017, dim...@ep... wrote:

> I would like to ask if libmesh can support conforming mesh adaptation.

Technically, "can support" - yes.  One of the consequences of the
"toolkit, not a framework" philosophy is that users can do a lot of
things without specific library APIs to do them, and looping through
elements to split some of them would be a good example.  libMesh will
even do a few things to help with that, like reconstructing neighbor
topology for you if you don't do that on the fly yourself.

But "can provide" - no.  Currently the only refinement types built
into the library are isotropic h refinement with hanging nodes,
isotropic p refinement on hierarchic element types, nodal
redistribution, and mixtures of those.

If you do write conforming mesh adaptation code, though, and it's
generic enough to work on others' problems (even if it's only for
triangles/tets), we'd love to incorporate it.
---
Roy

Re: [Libmesh-users] EquationSystems::write() / read() custom derived Systems

[Roy S. <roy...@ic...>]

On Mon, 2 Oct 2017, Salazar De Troya, Miguel wrote:

> If I have my custom derived System class, how can I save it to file
> with EquationSystems::write() and read it with
> EquationSystems::read(). It seems that they are not recognized and
> cast directly to their parents types.

Nope.  What you'll probably want to do is do the add_system() calls
yourself, then only do the EquationSystems::read() *afterwards*, as we
do in examples/adjoints/adjoints_ex3.  That way when the read needs
to fill your system, it will get the already-added system of the
correct type, and it won't have to figure out how to construct one
itself.

> I tried overriding System::system_type(), but now when reading, they
> get stuck in
> https://libmesh.github.io/doxygen/classlibMesh_1_1EquationSystems.html#a09450d5e4547d8b2e5690043da03dbd3,
> since that function cannot call custom derived Systems.

I'm afraid so.  I originally considered replacing this logic with a
factory method, but then the "add a custom system yourself then read
into it" idiom turned out to be so easy that I didn't think "make
user code construct its own factory map entry" would be an
improvement.

However, that idiom only works in apps that know what sort at compile
time what subclass of system they want to solve, and merely read in
data at run time.  If you're trying to write a more flexible run-time
tool, then you're out of luck unless you want to be the one who puts
factory support in add_system().
---
Roy

Re: [Libmesh-users] Conforming adaptivity

[John P. <jwp...@gm...>]

On Tue, Oct 3, 2017 at 2:33 AM, <dim...@ep...> wrote:

> Hello,
>
> I would like to ask if libmesh can support conforming mesh adaptation.
>

Not automatically, no.

There is definitely enough machinery in the library to allow conforming
adaptations/coarsenings of the mesh to be performed manually, but the
required code would probably be quite complicated...

-- 
John

[Libmesh-users] Conforming adaptivity

[<dim...@ep...>]

Hello,

I would like to ask if libmesh can support conforming mesh adaptation.

Thanks,

Dimitrios

[Libmesh-users] EquationSystems::write() / read() custom derived Systems

[Salazar De T. M. <sal...@ll...>]

Hi

If I have my custom derived System class, how can I save it to file with EquationSystems::write() and read it with EquationSystems::read(). It seems that they are not recognized and cast directly to their parents types. I tried overriding System::system_type(), but now when reading, they get stuck in https://libmesh.github.io/doxygen/classlibMesh_1_1EquationSystems.html#a09450d5e4547d8b2e5690043da03dbd3, since that function cannot call custom derived Systems.

Thanks
Miguel

--

Re: [Libmesh-users] Where to start?

[John P. <jwp...@gm...>]

On Sun, Oct 1, 2017 at 8:50 PM, Daniel Oliveira <dan...@ho...>
wrote:

> Hi friends,
>
> I am new on libmesh and I was thinking where I should start to read to
> understand the program.
> There is a book? Should I study using examples?
>

Unfortunately no, there's not a book per se, but you may want to peruse the
list of papers [0] that people have published as you may find one
approximately in the area of your research


My main objective is to develop structural transient analysis.
> Can someone point me the way?
>

The most relevant examples to study for structural analysis problems are:

systems_of_equations_ex4 -- 8
miscellaneous_ex11, ex12

[0]: http://libmesh.github.io/publications.html

-- 
John

Re: [Libmesh-users] Where to start?

[Tobias M. <tob...@un...>]

Hi Daniel,
I think, your idea to start with the examples is the best route.
At least I started by looking at the examples to see how the library is used. Maybe you can take one of them even as a starting point for your project.
I personally use the examples to find out, how certain features are meant to be used and use the documentation (http://libmesh.github.io/doxygen/annotated.html) to seek functions that fit my needs.
I hope, this helps a bit.
Best regards,
Tobias
________________________________________
Von: Daniel Oliveira [dan...@ho...]
Gesendet: Montag, 2. Oktober 2017 04:50
An: lib...@li...
Betreff: [Libmesh-users] Where to start?

Hi friends,

I am new on libmesh and I was thinking where I should start to read to understand the program.
There is a book? Should I study using examples?

My main objective is to develop structural transient analysis.
Can someone point me the way?

I thank you all in advance.

Daniel Oliveira
Whatsapp: +55 31 997 07 207
Pontíficia Universidade Católica de Minas Gerais - Brasil
------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
_______________________________________________
Libmesh-users mailing list
Lib...@li...
https://lists.sourceforge.net/lists/listinfo/libmesh-users

[Libmesh-users] Where to start?

[Daniel O. <dan...@ho...>]

Hi friends,

I am new on libmesh and I was thinking where I should start to read to understand the program.
There is a book? Should I study using examples?

My main objective is to develop structural transient analysis.
Can someone point me the way?

I thank you all in advance.

Daniel Oliveira
Whatsapp: +55 31 997 07 207
Pontíficia Universidade Católica de Minas Gerais - Brasil

Re: [Libmesh-users] high-order shape functions

[Roy S. <roy...@ic...>]

On Thu, 28 Sep 2017, Manav Bhatia wrote:

>  I noticed that Legendre functions are available in the code for
>  infinite FE approximations. I am curious if Legendre functions are
>  also available for DG solutions in libMesh.

I'm afraid not.

>  If not, how much effort would it be to migrate the infinite FE
>  implementation of Legendre polynomials for this purpose?

Implementing the 1D polynomials themselves isn't hard; the trouble is
getting the indexing right in 2D and 3D.  The MONOMIAL implementation
would be the closest to what you want, if you'd like to try adapting
it.
---
Roy

[Libmesh-users] high-order shape functions

[Manav B. <bha...@gm...>]

Hi, 

  I noticed that Legendre functions are available in the code for infinite FE approximations. I am curious if Legendre functions are also available for DG solutions in libMesh. 

  If not, how much effort would it be to migrate the infinite FE implementation of Legendre polynomials for this purpose?

Thanks,
Manav

[Libmesh-users] User's programing guide for the ceritfied reduced basis method with Libmesh

[Nguyen, A. <an...@ia...>]

Dear David,


I am writing you for your advice about applying the  "ceritfied reduced basis method" by you et al. with LIBMESH.


Is it possible to extend the open codes with LIBMESH for the shell elements?


I have read your related articles and really want to know how to program my problem based on your examples with LIBMESH.  Could you please give me some hints for getting started to program my defined problem.


Much thanks for your attention.  I look forward to hearing from you.


Best regards,


Dr.-Ing. An Danh Nguyen
Research Associate
==========================
Institute of General Mechanics
RWTH-Aachen
Templergraben 64
52062 Aachen, Germany
Tel: 0049 - 241 80 98282
Email:an...@ia...<mailto:Email:an...@ia...>

Re: [Libmesh-users] problems with running libmesh in parallel

[Roy S. <roy...@ic...>]

On Tue, 12 Sep 2017, Michael Povolotskyi wrote:

> Hi Roy, here is the progress.
> It turns out that openmpi 1.10.0 does not work as it should. Our admins tried 
> to rebuilt it with different options, but no success.
> So, we switched to openmpi 2.1.0 and it works.

Good to hear; thanks for the update!

I used to test libMesh with openmpi 1.10.0 occasionally... but looking
through module build logs, it seems I haven't even had a copy of
1.10.0 to test with in at least a year, possibly several.  It's quite
possible that we had (whether intentionally or not) a workaround for
some 1.10.0 bug at one point, then removed it without realizing.
---
Roy

Re: [Libmesh-users] problems with running libmesh in parallel

[Michael P. <mpo...@pu...>]

Hi Roy, here is the progress.
It turns out that openmpi 1.10.0 does not work as it should. Our admins 
tried to rebuilt it with different options, but no success.
So, we switched to openmpi 2.1.0 and it works.
Michael.

On 09/11/2017 10:42 AM, Roy Stogner wrote:
>
> On Mon, 11 Sep 2017, Michael Povolotskyi wrote:
>
>> This is exactly what had happened: the test runs till the end with 
>> MPICH and hangs here with openMPI.
>
> Great!
>
> (I realize now, after the fact, that I'd forgotten about an
> MPI_Get_count in the middle of all that; although I don't see how that
> would affect the problem I also don't *understand* the problem, so I'm
> glad my omission wasn't critical)
>
>> I'm going to inform our admins about this issue.
>
> Thanks; let me know how it goes!
>
> I'd still place my money on "something's wrong with that OpenMPI
> installation", but "Roy misunderstood something subtle about the MPI
> standard" is still in the running, so let me know if anyone ever
> figures out exactly what happened.  The MPI standard does have at
> least a few places where implementations are allowed-but-not-required
> to punish user mistakes, which can leave users with a program that's
> technically broken but works-for-me right up until it doesn't.
> ---
> Roy

Re: [Libmesh-users] Creating a parallel mesh with 2 cubes

[Roy S. <roy...@ic...>]

On Tue, 12 Sep 2017, Cyrill Vonplanta wrote:

> I am relatively new to libmesh and I would like to create a geometry
> in parallel with 2 cubes for benchmarking purposes. What would be
> the easiest way to go about this?
>
> I cannot call MeshTools::Generation::build_cube() 2 times with the
> same mesh, right?

I'm afraid not.  Not only does build_cube() clear any existing mesh
objects (and number newly created objects based on the assumption that
the clear occurred), but it does its new object creation in a
replicated fashion, relying on later prepare_for_use() to trim away
remote elements in a distributed mesh.  Fixing the first problem
probably wouldn't be too hard for you (it would just require making
use of a couple id offsets everywhere), but fixing the second would
basically require manually partitioning and ghosting the mesh on the
fly and would be quite tricky.  Usually when we want a cube-type
geometry in parallel we create a coarse version first (so the
replicated code is cheaper) and then uniformly refine it (which can be
done distributed-parallel).
---
Roy

[Libmesh-users] Creating a parallel mesh with 2 cubes

[Cyrill V. <cyr...@us...>]

Dear libmesh-users,

I am relatively new to libmesh and I would like to create a geometry in parallel with 2 cubes for benchmarking purposes. What would be the easiest way to go about this?

I cannot call MeshTools::Generation::build_cube() 2 times with the same mesh, right?

Best
Cyrill

Re: [Libmesh-users] problems with running libmesh in parallel

[Roy S. <roy...@ic...>]

On Mon, 11 Sep 2017, Michael Povolotskyi wrote:

> This is exactly what had happened: the test runs till the end with MPICH and 
> hangs here with openMPI.

Great!

(I realize now, after the fact, that I'd forgotten about an
MPI_Get_count in the middle of all that; although I don't see how that
would affect the problem I also don't *understand* the problem, so I'm
glad my omission wasn't critical)

> I'm going to inform our admins about this issue.

Thanks; let me know how it goes!

I'd still place my money on "something's wrong with that OpenMPI
installation", but "Roy misunderstood something subtle about the MPI
standard" is still in the running, so let me know if anyone ever
figures out exactly what happened.  The MPI standard does have at
least a few places where implementations are allowed-but-not-required
to punish user mistakes, which can leave users with a program that's
technically broken but works-for-me right up until it doesn't.
---
Roy