[CPMD-list] Platform-dependent output energies
Axel Kohlmeyer
axel.kohlmeyer at theochem.ruhr-uni-bochum.de
Sat Mar 27 07:35:57 CET 2004
On Fri, 26 Mar 2004, Eung-Gun Kim wrote:
EK> Dear CPMD Mailing List Subscribers,
EK>
EK> I have been running tests on different platforms and am getting different energy
EK> values on different machines. The platform-dependent results don’t show when I
EK> run on the sample input files that came with the code. With my own input file,
EK> however, I get values of the total energy as different as by 0.1 Ha upon
EK> switching the machine. The discrepancies exist before and after geometry
EK> optimization. Some relevant info on my input and installation is as follows:
EK>
EK> CP Version: 3.7.2
EK>
EK> Tested platforms:
EK> 1) IBM SP2 (AIX 5.2): xlf; MPICH 1.2.5.2; LAPACK 3.0
EK> 2) IBM Xeon cluster (RedHat 9.0): PGI Fortran compiler 5.1 (with BLAS and
EK> LAPACK); MPICH 1.2.5.2
dear Eung-Gun Kim and other cpmd with linux users,
please don't use the PGI provided BLAS and LAPACK libraries for any
serious calculation. they are horribly slow and produce incorrect results
with cpmd. and while i'm at it, please don't use the BLAS/LAPACK
libraries that linux distributions provide. they are usually compiled
with the gnu fortran compiler (g77) and thus depend on the g77 runtime
library which does not mix well with other fortran runtime libraries
(PGI is ok, IIRC, intel ifc has problems occasionally). on top of that,
they are usually compiled as shared libraries, which is a real
performance killer on the register starved 32-bit x86 platforms.
please use ATLAS or the intel MKL instead. depending on
the type of cpmd job, your calculations will run 1.8 to 2.5 times
as fast, while at the same time the results will be correct.
(since BLAS/LAPACK rountines account for only a part of the time
critical part of a cpmd run, the performance difference between
the libraries is much higher. from a test with the linpack benchmarks
i would estimate it as a factor of about 8).
ready to use, pgi compatible, and tested atlas libraries are at
http://www.theochem.ruhr-uni-bochum.de/~axel.kohlmeyer/cpmd-linux.html
please note, that even the generic (pentium-2) library will only
introduce a performance penalty of 5-10%, while using the
'standard' BLAS/LAPACK (if compiled properly) will give an overhead of
about 100-130%. the PGI compiled BLAS/LAPACK is even worse.
axel kohlmeyer.
EK>
EK> System:
EK> Aluminum crystal (2x2x2 cells for a total of 32 atoms)
EK>
EK> As you note in my input file below, I didn’t use either k-point sampling or FEMD
EK> for this metallic system. Portions from each output file are also attached at
EK> the end of the input file.
EK>
EK> Best regards,
EK>
EK> E Kim
EK>
EK>
EK> &CPMD
EK> OPTIMIZE GEOMETRY
EK> PCG MINIMIZE
EK> BFGS
EK> LSD
EK> EMASS
EK> 900
EK> TIMESTEP
EK> 5
EK> MAXSTEP
EK> 10000
EK> TRAJECTORY SAMPLE
EK> 10
EK> &END
EK>
EK> &DFT
EK> FUNCTIONAL BLYP
EK> &END
EK>
EK> &SYSTEM
EK> SYMMETRY
EK> 1
EK> ANGSTROM
EK> CELL ABSOLUTE
EK> 8.1 8.1 8.1 0.0 0.0 0.0
EK> CUTOFF
EK> 70
EK> CHARGE
EK> 0
EK> MULTIPLICITY
EK> 1
EK> &END
EK>
EK> &ATOMS
EK> *Al_MT_GIA_BLYP.psp KLEINMAN-BYLANDER RAGGIO=1.0
EK> LMAX=D
EK> 32
EK> 0.000000000 0.000000000 0.000000000
EK> 2.025000000 2.025000000 0.000000000
EK> 0.000000000 2.025000000 2.025000000
EK> 2.025000000 0.000000000 2.025000000
EK> 0.000000000 0.000000000 4.050000000
EK> 2.025000000 2.025000000 4.050000000
EK> 0.000000000 2.025000000 6.075000000
EK> 2.025000000 0.000000000 6.075000000
EK> 0.000000000 4.050000000 0.000000000
EK> 2.025000000 6.075000000 0.000000000
EK> 0.000000000 6.075000000 2.025000000
EK> 2.025000000 4.050000000 2.025000000
EK> 0.000000000 4.050000000 4.050000000
EK> 2.025000000 6.075000000 4.050000000
EK> 0.000000000 6.075000000 6.075000000
EK> 2.025000000 4.050000000 6.075000000
EK> 4.050000000 0.000000000 0.000000000
EK> 6.075000000 2.025000000 0.000000000
EK> 4.050000000 2.025000000 2.025000000
EK> 6.075000000 0.000000000 2.025000000
EK> 4.050000000 0.000000000 4.050000000
EK> 6.075000000 2.025000000 4.050000000
EK> 4.050000000 2.025000000 6.075000000
EK> 6.075000000 0.000000000 6.075000000
EK> 4.050000000 4.050000000 0.000000000
EK> 6.075000000 6.075000000 0.000000000
EK> 4.050000000 6.075000000 2.025000000
EK> 6.075000000 4.050000000 2.025000000
EK> 4.050000000 4.050000000 4.050000000
EK> 6.075000000 6.075000000 4.050000000
EK> 4.050000000 6.075000000 6.075000000
EK> 6.075000000 4.050000000 6.075000000
EK> &END
EK>
EK>
EK> Output from the SP2 run---------------------------------------------
EK>
EK> (K+E1+L+N+X) TOTAL ENERGY = -64.39023857 A.U.
EK> (K) KINETIC ENERGY = 28.11394543 A.U.
EK> (E1=A-S+R) ELECTROSTATIC ENERGY = -75.61888788 A.U.
EK> (S) ESELF = 114.89537676 A.U.
EK> (R) ESR = 0.00001993 A.U.
EK> (L) LOCAL PSEUDOPOTENTIAL ENERGY = -5.89039837 A.U.
EK> (N) N-L PSEUDOPOTENTIAL ENERGY = 13.16534369 A.U.
EK> (X) EXCHANGE-CORRELATION ENERGY = -24.16024144 A.U.
EK> GRADIENT CORRECTION ENERGY = -0.22825193 A.U.
EK>
EK>
EK> Output from the Xeon run--------------------------------------------
EK>
EK> (K+E1+L+N+X) TOTAL ENERGY = -64.27276166 A.U.
EK> (K) KINETIC ENERGY = 28.36257064 A.U.
EK> (E1=A-S+R) ELECTROSTATIC ENERGY = -76.82637865 A.U.
EK> (S) ESELF = 114.89537676 A.U.
EK> (R) ESR = 0.00001993 A.U.
EK> (L) LOCAL PSEUDOPOTENTIAL ENERGY = -6.46519951 A.U.
EK> (N) N-L PSEUDOPOTENTIAL ENERGY = 14.94091935 A.U.
EK> (X) EXCHANGE-CORRELATION ENERGY = -24.28467349 A.U.
EK> GRADIENT CORRECTION ENERGY = -0.29067887 A.U.
EK>
EK>
EK>
EK> _______________________________________________
EK> CPMD-list mailing list
EK> CPMD-list at cpmd.org
EK> http://www.cpmd.org/mailman/listinfo/cpmd-list
EK>
EK>
--
=======================================================================
Dr. Axel Kohlmeyer e-mail: axel.kohlmeyer at rub.de
Lehrstuhl fuer Theoretische Chemie Phone: ++49 (0)234/32-26673
Ruhr-Universitaet Bochum - NC 03/53 Fax: ++49 (0)234/32-14045
D-44780 Bochum http://www.theochem.ruhr-uni-bochum.de/~axel.kohlmeyer/
=======================================================================
More information about the CPMD-list
mailing list