RE: AMBER: Fwd: Amber9 parallel compilation openmpi issues

From: Ross Walker <ross.rosswalker.co.uk>
Date: Wed, 25 Jul 2007 08:25:58 -0700

Dear Francesco,

> Analysis of config.h and errors.log by comepetents in
> openmpi, who do not know
> Amber, may help. It is clear at this point that the issue can
> only be solved
> with help from Amber compilation experts. Notice please that

The issue here is that you are greatly over complicating things by trying to
resolve things in ten different directions at once, giving different error
messages to different people and generally just confusing everybody.

The key problem I believe is that you are receiving error messages early on
in the process that you are ignoring and proceeding to the next step
blindly. Then things are failing further down the line and it is at this
point that you post all the errors and people are left trying to work how on
earth this happened when all the previous steps (as you state) worked
without error. This results in people wasting a lot of time trying to work
out what is the bug in the Amber configure script that would cause this,
given that it works on a such large number of other machines. This just
results in everybody wasting time and the problem not getting solved.

The analogous experimental scenario would be designing a 10 stage synthesis
and not bother to to use NMR etc on each step to make sure it made what you
wanted. So step 2 fails and gives you a completely different product that
you expected but you carry on blindly. Then when you get to say stage 8 it
makes a ton of decomposed brown goo and this point you go and ask someone
why step 8 of your proposed synthesis doesn't work and this person then
spends ages trying to work such a simple step as step 8 doesn't work when
they have used it all the time in their work...

So lets take this in small steps one at a time and hopefully we can work out
where the problem occurs and fix it so that you can run Amber in parallel.
Firstly lets start with the mpi installation. I suspect that this has not
been correctly installed or has been installed to a directory other than you
think it has. Do you specifically know where you installed openMPI? From
what I can gather you are pointing MPI_HOME to the directory of the original
source code and not where you actually installed it. So I suspect the chain
of events is something like this:

1) Assume that openMPI installed correctly but with no idea what directory
it installed to.
2) Do not bother to check that openMPI test cases work.
3) Point MPI_HOME to the source directory for openMPI assuming that this is
where it installed.
4) execute amber's configure script - configure most likely gives a warning
about missing mpif.h etc etc - you ignore this warning and do not feel it
necessary to report it in your list of things that went wrong.
5) having ignored the warning you then attempt to build amber anyway and get
a whole load of unresolved dependencies and then post to the amber list
claiming that nothing works. \

The key point here is that the error could have been picked up at a number
of steps. At the openMPI build step by not running the tests there and or
not making sure you actually knew where things were installed.

At the configure stage when you would have received a warning.

Start by building openMPI giving it a prefix for the directory you want it
installed in. Also make sure that you will use the same compiler as you will
for Amber 9. I guess ifort. You can leave the c compiler as gcc it makes no
difference. Be sure to check that the openMPI configure script does indeed
use the compiler you chose.

Then once this is done and you have done make install go take a look in the
directory you gave with the prefix and make sure things like mpif90 are
present. See if you can build some simple example mpi programs - I assume
openMPI comes with a tests suite. Add the bin directory of this install
directory to your path and then logout and log back in and make sure that
>which mpirun / >which mpf90 etc are returning the path you expect.

You most likely have multiple MPI installations on your system and you need
to make sure you are actually using the one you want to use - don't just
assume it.

There will probably be a ton of MPI stuff from LAM MPI installed in /usr/bin
etc all built against g77 - Wow you just have to love the ANARCHY that is
Linux :-(... You need to make sure the path to your custom installed MPI is
before /usr/bin /bin /usr/lib etc etc so put it at the front of your path
variable and make sure when you execute mpif90 etc it does indeed use the
correct one.

If the above does not return the path where you installed openMPI (or says
command not found) then you need to fix this BEFORE attempting to build
amber.

Alternatively if you have admins for this machine who complain that openMPI
is in use by other applications etc and you cannot change it then get them
to install you a copy of openMPI on the machine that is built against ifort
and have them provide you with the necessary modifications to your
.bashrc/.cshrc etc in order to use this openMPI instead of the default one.
If they complain then just tell them "tough it is your job so get on with
it" - If they don't know how to do it then they are pretty useless admins
for a cluster so have them fired and find someone else who knows what they
are doing.

Once you have sorted this and have a working MPI installation with a mpif90
--showme that reports the correct fortran compiler and you know what to set
MPI_HOME to then proceed to building Amber. First run configure and check
for any and all warnings / errors and act on them. If you don't understand
them then STOP at this point and ask the mailing list.

Then go ahead and build AMBER and hopefully this time it will work. If not
then before you modify anything / tweak paths etc then post to the AMBER
list with the errors and details about what you did - what configure
reported, what arguments you gave configure etc. Running "env" to show all
your environment variables can also help.

All the best
Ross

/\
\/
|\oss Walker

| HPC Consultant and Staff Scientist |
| San Diego Supercomputer Center |
| Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
| http://www.rosswalker.co.uk | PGP Key available on request |

Note: Electronic Mail is not secure, has no guarantee of delivery, may not
be read every day, and should not be used for urgent or sensitive issues.


-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Sun Jul 29 2007 - 06:07:12 PDT
Custom Search