Thanks for your help.
We have installed ifort instead of gfortran, and the speed of MEAM computation increases in ~ 10 times.
Best Regards.
Alexander
Dear Prokhoda
I have installed ICC for compiling lammps and I have used gfortran (default for Linux) to compile
and build libmeam.a to enable MEAM.
1-you mean, it is better to install ifort in linux then compile the MEAM and
build libmeam.a by ifort and then compile lammps by ICC?
2- is there conflict between gfortran and ifort when both of them have been installed?
3- is there conflict between icc and ifort when both of them have been installed?
Thanks
Hadi
Dear Prokhoda
I have installed ICC for compiling lammps and I have used gfortran (default
for Linux) to compile
i've made quite a few tests and in most cases there is next to no
difference in the LAMMPS performance for current hardware between
recent versions on intel icc and gnu gcc. an example of gcc compiler
flags for good optimization is here. you can try using -O3 instead of
-O2, which is sometimes faster, sometimes slower.
http://git.icms.temple.edu/git/?p=lammps-icms.git;a=blob;f=src/MAKE/Makefile.openmpi-omp;hb=HEAD
and build libmeam.a to enable MEAM.
1-you mean, it is better to install ifort in linux then compile the MEAM and
build libmeam.a by ifort and then compile lammps by ICC?
libmeam is written in fortran and unlike icc/gcc the intel fortran
compiler does generate a bit faster code. however, the makefile for
gfortran in lib/meam only uses a moderate optimization. raising the
optimization level may result in much faster code generated by
gfortran. the large reported difference in performance seems bogus to
me. this is likely due to some other interactions and due to incorrect
testing (e.g. of too small/short) a test run.
2- is there conflict between gfortran and ifort when both of them have been
installed?
no. not if you do it correctly. the is one important issue: you have
to compile *all* of the fortran code included in lammps with the same
fortran compiler.
3- is there conflict between icc and ifort when both of them have been
installed?
no. why? they are provided by the same vendor.
in general, i would be careful with performance claims that have no
solid evidence accompanying them. there are far too many ways to make
mistakes. good performance measurement requires a lot of care and
almost anal-rententive attention to detail.
axel.