[mvapich-discuss] 32-bit MVAPICH and dynamic make scripts
Adam Moody
moody20 at llnl.gov
Mon Jun 26 18:23:40 EDT 2006
Hello Panda et. all,
We're planning to support a 32-bit environment on our x86-64 AMD systems
to provide a workable system while users port to 64-bit. Are there any
known problems in either MVAPICH or the OpenIB layer when building and
running in 32-bit mode?
Also, we support many (currently 4) different compilers on our systems
and each must link with MPI. We'd like to limit the number of MPI
builds. While most compilers build compatible binaries from C source
code, there are still some incompatibilities between compilers with C++
and Fortran code. Specifically, it's often necessary to at least build
compiler-specific libraries for libmpi++ and libfarg. This brings up a
couple questions / feature requests:
1) A make script which can be dynamically configured would help quite a
bit. For example, it would be nice to be able to do the following:
export CC=gcc
export F77=g77
export PREFIX=/usr/lib/mpi/gnu
make.mvapich.gen2
export CC=icc
export F77=ifort
export PREFIX=/usr/lib/mpi/intel
make.mvapich.gen2
Currently, these variables are hard coded in make.mvapich.gen2, so that
one must edit the file manually or repeatedly apply patches, which is a
bit clumbsy. I think you could achieve the above by conditionally
setting the variables in make.mvapich.gen2 like so:
CC=${CC:-gcc}
This sets CC to the value of $CC if it is set and uses "gcc" otherwise.
You could apply this to other parameters users may like to customize
such as IBHOME, OPT, etc. An alternative would be to process
command-line arguments.
Additionally, it would be helpful to be have a variable the user could
set to specify the PCI device and speed in order to avoid the
interactive prompt.
2) In some cases, I think we could get by with a single build of the
MPI body, so long as we provide compiler-specific pieces like libmpi++
and libfarg. Would it be possible to set the build script up to assist
with just building these pieces and not the whole MPI?
Thanks,
-Adam
========================
Adam Moody
Development Environment Group
Lawrence Livermore National Laboratory
More information about the mvapich-discuss
mailing list