[maker-devel] WARNING: Multiple MAKER processes...

[Carson Holt]

Hi Jenny,

There are three locations that are key to running with MPI.

1. The location of mpicc
2. the location of mpi.h
3. the location of mpiexec

These files are common to all MPI flavors, but are incompatible between
MPI flavor and version.  For example, if you use mpicc to compile with the
wrong mpi.h, it will fail. If you compile and then try and execute with
the wrong mpiexec, it will fail.  So you need to make sure all 3 are
coming from the same version when running maker or any MPI program.  A
common mistake is to just call mpiexec direct on the command line.  Your
system may in fact call the wrong mpiexec when you do that (use ‘which
mpiexec’ to see which one is being called when you just type mpiexec).
Also when configuring MAKER, manually review the path to mpicc and mpi.h
given during the ‘perl Build.PL’ configuration step to make sure they are
coming from the same version and flavor of MPI.

Also there is no mpi_maker. The standard 'maker' command will now be MPI
enabled if you say ‘yes' to the 'configure for MPI' question during the
'perl Build.PL' step of the install.

I just installed MPICH3 and ran MAKER on a test job.  Here is what I did
step-by-step —> 

#first install MPICH3
wget http://www.mpich.org/static/downloads/3.0.4/mpich-3.0.4.tar.gz
tar -zxvf mpich-3.0.4.tar.gz
cd mpich-3.0.4

#note that I configure with shared libraries enabled *this is not the
default!!
#also set prefix to wherever you want it installed
./configure --enable-shared=yes --prefix=/usr/local/mpich3
make
make install

#now cofigure MAKER
cd /usr/local/maker/src
perl Build.PL
#say yes to MPI question
#give full path to mpicc —> /usr/local/mpich3/bin/mpicc
#give full path to mpi.h containing directory —> /usr/local/mpich3/include
./Build
./Build install

#test using full paths to executables
/usr/local/mpich3/bin/mpiexec -n 3 /usr/local/maker/bin/maker



In addition, use MAKER 2.30.  It has the best MPI performance of any
version of MAKER.  Also you can use OpenMPI rather than MPICH if you
choose.  It actually scales better than MPICH. The main difference is that
you must set LD_PRELOAD to the location of libmpi.so before even trying to
install MAKER. It must also be set before running MAKER (or any program
that uses OpenMPI's shared libraries), so it's best just to add it to your
~/.bash_profile. (i.e. export
LD_PRELOAD=/usr/local/openmpi/lib/libmpi.so).  Note however that having
this set will cause errors for MPICH, so be careful if switching back and
forth between MPI flavors.


Thanks,
Carson

 

On 1/28/14, 5:14 AM, "Jenny Barna" <jcjb at cam.ac.uk> wrote:

>
>We have had maker since 2010. Maker 2.10 and 2.11 compiled with mpich2
>were and still are working with mpi as far as I can tell. These have
>a binary called mpi_maker.
>
>I have kept the same mpich2 installed to try to make sure things stay
>working. It has been picked up by subsequent maker builds.
>It is mpich2-1.2.1p1 .
>
>I have seen the posting by Marc P. Hoeppner on 10/26/13 about
>'WARNING: Multiple MAKER processes have been started in the same
>directory.' which is the same warning I have had.
>
>Trying all my previous installations I now see this is happening in all
>2.2? distributions I have. Versions 2.10/2.11 are not doing this.
>It is happening with 2.2?/2.30p where no separate mpi_maker binary is
>made.
>
>Trying new mpich 3.0.4 as advised in the posting by Marc has not worked.
>It appears to build but fails at once with severe mpi errors. Tests
>included
>with that mpi distro worked.
>
>I am aware the maker WARNING behaviour is indicative of mpi types having
>become mixed but have not been able to correct.
>
>
>
>Jenny Barna                              | jcjb at cam.ac.uk
>Bioinformatics and Computational Biology | computing.bio.cam.ac.uk
>Dept of Biochemistry                     | +44 1223 333644
>80 Tennis Court Road, Cambridge, CB2 1GA, UK
>
>
>_______________________________________________
>maker-devel mailing list
>maker-devel at box290.bluehost.com
>http://box290.bluehost.com/mailman/listinfo/maker-devel_yandell-lab.org