My Community

Hi,

i have a minor problem with vasp.5.2.12/gfortran 4.1.2:

./preprocess <bse.F | /usr/bin/cpp -P -C -traditional >bse.f90 -DHOST=\"Linuxgfortran\" -Dkind8 -DCACHE_SIZE=12000 -Dgfortran -Davoidalloc -DNGXhalf
gfortran -ffree-form -ffree-line-length-none -I/opt/fftw3/3.2.1/include -O3 -funroll-all-loops -c bse.f90
In file bse.f90:153

INTEGER, ALLOCATABLE :: N1(:) ! valence band index
1
Error: Attribute at (1) is not allowed in a TYPE definition

Similar for:
INTEGER, ALLOCATABLE :: N3(:) ! conduction band index
INTEGER, ALLOCATABLE :: NK(:) ! k-point index
INTEGER, ALLOCATABLE :: ISP(:) ! spin index
INTEGER, ALLOCATABLE :: INDEX(:,:,:,:) ! yield index into Hamilton matrix

It seems that gfortran 4.1.2 is old, but this can be fixed by using pointer instead of allocatable,
see http://www.groupsrv.com/computers/about548043.html and http://code.google.com/p/fds-smv/issues/detail?id=1001.
Fix would be valuable as gfortran 4.1.2 is still the main compiler on RedHat 5, which is probably the most common system on custom beowulf clusters.

I am experiencing exactly the same problem with VASP 5.2.12 on a CentOS 5 cluster. GCC 4.6.1 yields a faulty executable (BRMIX: very serious problems...) and lowering the optimization for broyden.F, even to -O0 did not help. GCC 4.4.4 fails to compile at some point.

For me, only GCC 4.3.3 produces a working (OpenMPI) executable (with gotoblas 1.06, fftw3.3). Although it's an Opteron machine, ACML is significantly slower. Atlas was not tested.

I compared my results to those obtained with Intel 11.1.069/mkl/fftwfurth (makefile.linux_ifc_P4). I obtain the exact same forces and energy, which is good, and the gcc executable is ca. 20 % faster. An optimized (-O3 -msse3) intel12.0.3/mkl/fftw3.3 executable yields again the same results, but gcc is 7 % faster. I am a little surprised by the poor performance of the intel compiler.

By the way, I was able to compile 4.6.38 with gcc 4.1.2/gotoblas2 1.06/fftw3.3. The resulting executable produces again identical results to intel 11.1 and is significantly faster.



<span class='smallblacktext'>[ Edited Sun Oct 09 2011, 03:54AM ]</span>

Could the difference come mostly from fftwfurth vs fftw3?
Make also sure that intel mkl and goto do not use threads (ps -fL).
Have you tried running benchmarks?
http://cms.mpi.univie.ac.at/vasp-forum/ ... php?2.9576
I imagine this is a set of regressions tests, but don't have access to it yet.
I'm surprised vasp tarball does not contain regression tests.