Thanks! I built against a locally built valgrind and ran valgrind on a mpi program, there's a lot of noise, but hopefully it will be useful to mpi developers here.

Related to building a mpi debug package, is there a simple example of using vampirtrace? I've

* found documentation on running the vt compiler wrappers
* made a small example that generated a .otf, .thumb and set of .events.z files (the events file are one per rank?)
* found the vampirtrace pdf in the installation

But it sounds like I am supposed to run a GUI, vamptrace, to interpret the results? This doesn't appear to be part of the open-mpi installation. How would I get this and see the communication in my simple example?

I'm also wondering, since our developers are mostly going to use python with mpi4py, if I should make a build of mpi4py with the vt wrappers.

best,

David

On 07/14/16 08:44, Jeff Squyres (jsquyres) wrote:
The key is that you need to specify --with-valgrind=valgrind_install_dir -- not 
the path to the valgrind executable.

Additionally, there's a valgrind.h that you'll need to have in that tree.  
E.g., if you specify --with-valgrind=/opt/valgrind, then it expects to find 
/opt/valgrind/include/valgrind/valgrind.h in that tree.

If your system doesn't have that, you might need to look for a "dev" or "devel" kind of 
valgrind package to supplement your main "valgrind" package, and/or install Valgrind from source.

Hope that helps.


On Jul 14, 2016, at 11:32 AM, David A. Schneider <david...@slac.stanford.edu> 
wrote:

I thought it would be a good idea to build a debugging version of openmpi 1.10.3. 
Following the instructions in the FAQ: 
https://www.open-mpi.org/faq/?category=debugging#memchecker_how 
<https://email.slac.stanford.edu/owa/redir.aspx?REF=RP5_rOiQ9WZQgiVLWWUfJLansCsxNMkiYakBQ0zCua7x_z9A-6vTCAFodHRwczovL3d3dy5vcGVuLW1waS5vcmcvZmFxLz9jYXRlZ29yeT1kZWJ1Z2dpbmcjbWVtY2hlY2tlcl9ob3c.>
 I went with

.configure —with-verbs —enable-debug —enable-memchecker 
—with-valgrind=/usr/bin/valgrind

Since I discovered valgrind was already installed on our network (this is 
rhel7). However configure fails, and part of the error messages read

WARNING Expected file /usr/bin/include/valgrind/valgrind.h not found

Is there a way to build openmpi to use the installation of valgrind on my 
system? I’m not sure if/where the valgrind.h file is, I don’t see it in 
/usr/include. If not, I will build against my own build of valgrind, but I run 
into similiar error messages - the auto tools script is not finding the headers 
my valgrind install - I've tried giving a full path to valgrind, the directory 
I used for PREFIX when I installed valgrind, there must be some assumptions 
about environment variables I need to have set to get the openmpi install 
script to work?

Reading through a page on valgrind: 
http://valgrind.org/docs/manual/mc-manual.html#mc-manual.mpiwrap 
<https://email.slac.stanford.edu/owa/redir.aspx?REF=RUoQIaKZhwLZN3KFa0DpMOj3B1mLOQzP7xrqLqeC7obx_z9A-6vTCAFodHRwOi8vdmFsZ3JpbmQub3JnL2RvY3MvbWFudWFsL21jLW1hbnVhbC5odG1sI21jLW1hbnVhbC5tcGl3cmFw>

I found instructions to the effect of how to wrap valgrind around an existing 
open-mpi install. Does any one know which is better? That is, what I'm trying, 
to build open-mpi based on a valgrind install, or to install valgrind based on 
a open-mpi install? Doing both looks like a circular dependency.

Best,

David
_______________________________________________
users mailing list
us...@open-mpi.org
Subscription: https://www.open-mpi.org/mailman/listinfo.cgi/users
Link to this post: 
http://www.open-mpi.org/community/lists/users/2016/07/29668.php


Reply via email to