optimization failure?

Message

MYoshiya · #1 Post by **MYoshiya** » Tue Nov 20, 2007 10:07 am

Hi all.

I'm currently using ifort ver. 10.0.023 with mpich-1.2.6 or mpich2-1.0.5p4 for parallel execution. I had no problem before I upgraded ifort from ver. 9.X with the same versions of mpich and mpich2, but with ver. 10.0.023 or 10.1.008 which is currently available, executing parallel VASP for bench.Hg resulted in serious warnings shown at the bottom.

Decreasing optimization level in makefile to -O1 solved the problem, which implies this is caused by optimization upon compilation. But serial execution showed no problem at all without changing optimization level, which means compiler itself is okay. mpich or mpich2 had been no problem with earlier version, which means configurations of mpichX is okay. What is responsible for the serious warning? Does anybody has experienced similar thins to this? Or are there any negative information for combination of VASP-ifort-10?

Any info is appreciated.

Thank you very much in advance.

------- Summary of warning in bench.Hg calc.
N E dE d eps ncg rms rms(c)
RMM: 1 -0.519878279328E+05 -0.51988E+05 -0.13845E+05 316 0.792E+02
RMM: 2 -0.529309060433E+05 -0.94308E+03 -0.24379E+04 316 0.246E+02
RMM: 3 -0.530742153645E+05 -0.14331E+03 -0.42491E+03 316 0.126E+02
RMM: 4 -0.532159680227E+05 -0.14175E+03 -0.15817E+03 316 0.800E+01
RMM: 5 -0.532810234263E+05 -0.65055E+02 -0.69776E+02 316 0.474E+01
RMM: 6 -0.533391932197E+05 -0.58170E+02 -0.58908E+02 748 0.328E+01
RMM: 7 -0.533471687475E+05 -0.79755E+01 -0.13912E+02 793 0.968E+00BRMIX: very serious problems
the old and the new charge density differ
old charge density: 597.07135 new 595.99206
0.238E+01
RMM: 8 -0.532559068168E+05 0.91262E+02 -0.33250E+02 680 0.178E+01BRMIX: very serious problems
the old and the new charge density differ
old charge density: 596.10337 new 596.46068
0.942E+00

job · #2 Post by **job** » Tue Nov 20, 2007 12:19 pm

While I haven't experienced the problems you mention, the following assumption

But serial execution showed no problem at all without changing optimization level, which means compiler itself is okay.

cannot be relied upon. It perfectly possible for the compiler to do some aggressive optimizations that happen to produce incorrect results only with the MPI version.

MYoshiya · #3 Post by **MYoshiya** » Tue Nov 20, 2007 11:41 pm

Thanks for reply. I agree with you, in that not showing warning in a calculation does not mean any other calculation is safe. At the moment, I decrease optimization level even for serial execution. But I'm not sure what is the real origin of the warning and why it hasn't fixed yet in newer version of Intel compiler (ver. 10.1.X). I suppose many VASP users use Intel Fortran compiler, don't they? Please share your experience with me.

#4 Post by **admin** » Tue Nov 27, 2007 11:44 am

version 10.1.008 is the most recent version that we have installed, this one works ok (with LAM for parallelization). As we have skipped the 10.0 releases, I cannot tell whether this compiler version worked ok to compile vasp.

MYoshiya · #5 Post by **MYoshiya** » Thu Nov 29, 2007 1:02 pm

Thank you very much for your comment. Using that version didn't solve my problem in combination with mpich-1.2.6 or mpich2-1.0.5p4 as someone else reported in the new thread. I'm trying to compile it under 64bit enviroment. Maybe the emt64 version of the compiler caused the problem, or perhaps, I need to switch to LAM.
Thank you, anyway.

#6 Post by **admin** » Fri Nov 30, 2007 9:15 am

please note that in mpich all integers are defined in C-standard (32bit), maybe there was a conflicting declaration of integers in your makefile?

MYoshiya · #7 Post by **MYoshiya** » Tue Dec 25, 2007 10:21 am

Output of ./configure of mpich2-1.0.5p4 showed
#define SIZEOF_F77_INTEGER 4
which seems to me okay.

By the way, today I tried openmpi-1.2.4 (a successor of LAM) with those intel compilers and results were the same: showed wrong results if optimization level didn't decreased to 1.

Humm, something is wrong but I don't know.

ravi · #8 Post by **ravi** » Thu Jan 17, 2008 1:48 pm

Hello,

I am also trying to install vasp4631 with /opt/intel/mkl/10.0.011/lib/em64t . Though vasp compiled without any error message the test calculation failed. I have tried for the fcc Cu for that I am getting the following message

##
running on 2 nodes
distr: one band on 1 nodes, 2 groups
vasp.4.6.31 08Feb07 complex
POSCAR found : 1 types and 1 ions
LDA part: xc-table for Ceperly-Alder, standard interpolation
POSCAR, INCAR and KPOINTS ok, starting setup
WARNING: wrap around errors must be expected
FFT: planning ... 1
reading WAVECAR
WARNING: random wavefunctions but no delay for mixing, default for NELMDL
entering main loop
N E dE d eps ncg rms rms(c)
DAV: 1 0.282009941497E+02 0.28201E+02 -0.64019E+03 1698 0.177E+03
DAV: 2 -0.947871324608E+01 -0.37680E+02 -0.35331E+02 1688 0.194E+02
DAV: 3 -0.101411396801E+02 -0.66243E+00 -0.66004E+00 1896 0.311E+01
DAV: 4 -0.101425808655E+02 -0.14412E-02 -0.14411E-02 1820 0.136E+00
DAV: 5 -0.101425829177E+02 -0.20522E-05 -0.20523E-05 1794 0.495E-02BRMIX: very serious problems
the old and the new charge density differ
old charge density: 11.36486 new 11.30349
0.921E+00
DAV: 6 -0.527231724835E+01 0.48703E+01 -0.27211E+01 1862 0.863E+01BRMIX: very serious problems
the old and the new charge density differ
old charge density: 11.31561 new 11.35910
0.507E+00
DAV: 7 -0.419205336849E+01 0.10803E+01 -0.42166E+00 1860 0.342E+01BRMIX: very serious problems
the old and the new charge density differ
old charge density: 11.35003 new 11.33817
0.916E-01
##
So, I am unable to converge the electronic minimization even for 60 scf cycle.

I hope somebody may have working version of makefile for mkl/10.0.011.

#9 Post by **admin** » Wed Jan 23, 2008 2:48 pm

do you get reasonable results if you run vasp serial?