Top Page > Browsing

Differing results from openmx -runtest for gcc and ifort Date: 2008/02/08 23:51 Name: J. J. Ramsey   <jjr19@uakron.edu> When doing the automatic test calculations "openmx -runtest" using ifort 10.1 (20071116) and an installation of MPICH2 compiled using icc and ifort, the results of the 13th test looked suspicious. For CC = mpicc -O3 and LIB = -lmkl_lapack -lmkl -lguide -lpthread -lfftw3, I get this: 13 input_example/Mol_MnO.dat Elapsed time(s)= 58.90 diff Utot= 0.006161915946 diff Force= 0.086939420494 For CC = mpicc -O and the same value of LIB as before, I get this: 13 input_example/Mol_MnO.dat Elapsed time(s)= 62.90 diff Utot= 0.006161915946 diff Force= 0.086939420494 When doing the automatic test calculations "openmx -runtest" using gcc 4.1 and an installation of MPICH2 compiled using this gcc and gfortran 4.2 (both part of my Ubuntu 7.10 distribution), the results seemed much saner. Here CC = mpicc -O3 and LIB = -L/usr/lib/atlas -llapack -lblas -lg2c -lfftw3. This is the result for the 13th test: 13 input_example/Mol_MnO.dat Elapsed time(s)= 134.69 diff Utot= 0.000000004967 diff Force= 0.000000000141 I noticed not only that the diffs are smaller, but the elapsed time is much larger than it is for the results using the Intel compilers. I couldn't find a sign, though that the 13th test under the Intel compilers died prematurely, though maybe I didn't know what to look for. Any ideas on what is going on?

Page: [1]

Re: Differing results from openmx -runtest for gcc and ifort ( No.1 ) Date: 2008/02/22 23:10 Name: Zsolt I also reported a similar problem earlier, here: http://www.openmx-square.org/forum/patio.cgi?mode=view&no=655 Although still no clue, what the problem was. It was compiled on AIX with mpicc. Zsolt Re: Differing results from openmx -runtest for gcc and ifort ( No.2 ) Date: 2008/02/28 07:04 Name: T.Ozaki Hi, This kind of problem is highly depending on computational environment including processor, OS, library, and compiler. So, I just list my experiences below: (1) The compiler option "-O3" sometimes gave wrong results. In this case, "-O1" or "-O2" solved the problem. (2) MKL sometimes gave wrong results. In this case, ATLAS solved the problem. If you post your experiences on above mentioned tips, it would be helpful for other users. Regards, TO Re: Differing results from openmx -runtest for gcc and ifort ( No.4 ) Date: 2009/12/08 12:01 Name: N.Kolchenko   <nkolchenko@mail.ru> Hi, Results for file Mol_MnO.dat with different libraries, platforms and XC-functionals(!). "Home" computer - Fedora-11, compiler - gcc 4.4, CCflag = -O3 -fopenmp -Dnompi -mtune=native. "Cluster" - CentOS, compiler - mpicc, CCflag = -O3 openmp. 1. NO DIFFERENCE with different optimization levels (-O2 or -O3). 2. NO DIFFERENCE with different libraries (MKL or ATLAS). 3. NO DIFFERENCE with different platforms ("home" or "cluster"), if you didn't use GGA-PBE. In the GGA-PBE mode results for "home" (OpenMP only) and "cluster" (MPI + OpenMP) are different (for above mentioned Mol_MnO.dat - ABS(Utot(cluster)-Utot(home)) ~ 0.006 a.u., in agreement with yours data). For LSDA-CA functional (for example) - everything OK. The same data can be easy obtained for another input files (Met.dat, for example). Conclusion - something may be wrong with MPI implementation in GGA mode (for "OpenMP only" the results are in agreement with ../input_example/Mol_MnO.out). Regards, NK P.S. Of course elapsed times are different for MKL and ATLAS, but slightly. May be ATLAS was not tuned sufficiently for your platform (but 135s/63s is too large value...). Check also: 1. Number of SCF steps in SCF history; 2. Number of threads in OpenMX launch file. Re: Differing results from openmx -runtest for gcc and ifort ( No.5 ) Date: 2010/02/03 15:52 Name: Kim, Minsung  <mskim1729@gmail.com> I had the same problem with the intel compiler and ACML(also mkl, ATLAS) in AMD Opteron cluster. But, I solved it by using -O1 option.

Page: [1]