Compiling MPICH2 with 7.0 fails at very end of installation

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

I’m trying to compile mpich2-1.0.5p4
On “make install”, I get:
…
pgcc -shared -fpic -DPIC .libs/trace_input.o .libs/rlogutil.o
.libs/logformat_trace_InputLog.o -lc -Wl,-soname -Wl,libTraceInput.so
-Wl,-version-script -Wl,.libs/libTraceInput.ver -o .libs/libTraceInput.so
/usr/bin/ld: .libs/trace_input.o: relocation R_X86_64_PC32 against `__pgio_ini’ can
not be used when making a shared object; recompile with -fPIC
/usr/bin/ld: final link failed: Bad value
make[2]: *** [INSTALL_LIBSO] Error 2
…

Current flags I’m using:
export=CFLAGS=“-fast "
export=CXXFLAGS=”-fast -fPIC -DPIC"
export=FFLAGS=“-fast "
export=F90FLAGS=”-fast "
export=LDFLAGS=“-fast "
export=OPTFLAGS=”-fast "
export FC=“pgf77”

What am I doing wrong? I see http://www.pgroup.com/userforum/viewtopic.php?p=3277&sid=e5f50252d1580fdb6c7ec6280da2282f
had something sort of similar, but I don’t think I’m using torque. Are there other flags I should be using?

I’m running this on Rocks 4.2.1 which is a modified CentOS 4.4.

Thanks for any advice!
Alex

2

Alex,

Are these 32-bit compilers or 64-bit compilers you are using?
Did you follow our tips and technique for MPICH2?
You will first need to update your export command from export=CFLAGS=“-fast” to export CFLAGS=“-fast” and so on.

Here are commands you can use on csh, assuming you install compiler in /opt/pgi.

% setenv PATH /opt/pgi/linux86-64/7.0/bin:$PATH
% env CFLAGS=“-fast " CXXFLAGS=”-fast" FFLAGS=“-fast " F90FLAGS=”-fast " LDFLAGS=“-fast " OPTFLAGS=”-fast " CC=“pgcc” CXX=“pgCC” F90=“pgf90” FC=“pgf77” CPP=“pgCC -E” ./configure --prefix=/path/to/install/dir

Let me know if you still have a problem.

If this is the only error you have, you can ignore unless you are planning to use MPI trace.

Hongyon

3

I don’t know how the extra = got in there! I wasn’t using those.

I’m following your MPICH2 stuff, and compiling with the 64bit compiler for Opterons.

Is mpi-trace important?

Thanks!
Alex

4

Alex,

You can do without MPE. I believe MPE is used for profilng MPI.

If you just want to run MPI program, mpi-trace is not needed. In fact, you can try by adding --without-mpe when you do configuration.

We just updated our cluster to ROCKS 4.3 and PGI 7.0-5 compilers. We didn’t see similar problem.

We will continue to investigate and will let you know if we find similar error.

Hongyon

5

Do you have any Rocks support documentation somewhere? I’m using 4.2.1. Perhaps I have misconfigured/broken something?

6

Alex,

You can find Rocks documentation at http://www.rocksclusters.org/rocks-documentation/4.2/.

Do you have a problem building a shared library with gcc?

Hongyon