vasp.5.2.2 - WARNING in EDDRMM: call to ZHEGV failed

[bademoor]

Dear VASP users,

I recently compiled the newest vasp.5.2.2 version without too much problems using following versions of compilers / libraries / ...
- intel fortran compiler 11.1.038
- intel math kernel library 10.2.0.013
- intel mpi 3.2.1.009
- intel fftw3 interface

I was able to compile AND to run the vasp.4.6 version using these tools (just to inform you that it works with these tools).

For the vasp.5.2.2, I can run my own "benchmarks" without any problem, I also find the same results (that is at least OK), however I get (several) warnings in the stdout, like:

Code: Select all

WARNING in EDDRMM: call to ZHEGV failed, returncode = 

followed by some number which are always different. So, I started searching on the forum and I found some other people also had these warning, however always with vasp.4.6, for which I never encountered suchlike warnings.
The most interesting among these forum-posts in my opinion was http://cms.mpi.univie.ac.at/vasp-forum/ ... .php?3.214, since some suggestions have been made there by the admin. I carefully cheched / tried these and I discuss my findings shortly below:

1. using ALGO=Fast (instead of IALGO=48) reduces the number of warnings, but does not exclude them. Using ALGO=Normal no warnings are printed to stdout, however since the blocked Davidson (DAV) is slower than the RMM-DIIS algorithm, I prefer not to use the ALGO=normal tag as standard setting in my INCAR file
2. input geometry is reasonable, as it is optimized in a few steps. With regard to the IBRION tag, using the Quasi Newton instead of the Conjugate gradient algorithm reduces the number of warnings, but still some appear.
3. Lapack was certainly correctly installed since everything worked properly with vasp.4.6. Moreover, the same warnings were printed when using use the LAPACK which is delivered with the code (vasp.5.lib/lapack_double.o)
4. Commenting the line "#define USE_ZHEEVX" and recompiling was not necessary since vasp did not produce warning if using the blocked Davidson (DAV) algorithm.

Another important remark: apparently, the warnings only appear in the gamma-only version and NOT in the complex version of vasp.5.2.2


MY QUESTIONS:
1. do I have to worry about the warnings?
2. is there a way to get rid of the warnings?
3. why only the gamma-only version produces these warnings and not the complex version of vasp.5.2.2?
4. Did any other vasp.5.2.2 users encountered the same problems? Any tips?


Hoping for a clarifying answer,
With best regards,

Bart

[bademoor]

Hi,
Is is possible for (one of) the admins to comment on this? I'm really wondering whether I have to worry about the warnings or whether I did something wrong.
Thx in advance
Bart

[admin]

sorry, I cannot reproduce this behaviour with the tests of the gamma-only version I ran, neither with the serial nor with the parallel executable
(Intel compiler 9.5,
openmpi-1.2.6,
GOTO's BLAS,
vasp's lapack,
FFT3D: both, Furthmueller's fft or fftw-3.1.2 of fftw.org),
therefore it's hard to tell

[bademoor]

Dear admin,

thx for the answer.
I'll "ignore" the warnings right now since the results are not affected (when I compare vasp.5.2 and vasp.4.6). I'll look further into this in future (update or other compilers / libraries and recompile).
When I would be able to get rid of the warning, I'll certainly post it here on the forum.

Regards
Bart

[martonak]

I compiled vasp.5.2.2 with ifort 11.1.064, openmpi-1.4, acml4.2.0 and fftw-3.1.2 and started testing it on some systems where vasp 4.6 runs without any problems. I observe almost all the time the warning

WARNING in EDDRMM: call to ZHEGV failed, returncode = 6 3 **

Just as reported in the original posting, I see these warnings only with the gamma point version, not with the complex one.

Interestingly, with NSIM=8 the gamma point version also crashes, depending on the NPAR value, e.g. running on 16 cores it runs for NPAR=2 (with the above warnings) but crashes for NPAR=4 with the message

mpirun noticed that process rank 12 with PID 4461 on node compute-0-3 exited on
signal 11 (Segmentation fault).

With NSIM=4 it does not crash.

I never observed such behaviour with various versions of VASP 4.6 ? Is there any solution to these problems ?

Many thanks in advance for suggestions.

Best regards

Roman Martonak

[fish]

I have observed the same message

"WARNING in EDDRMM: call to ZHEGV failed, returncode ="

after building the gamma point version of vasp.5.2.2 using the following versions of compilers and libraries

-intel fortran compiler 11.1.069
-intel math kernel library 10.2.2.025
-mvapich2 1.4.1
-intel fftw3 interface

on a Intel(R) Xeon(R) CPU X5550 @ 2.67GHz running Red Hat Enterprise Linux Server release 5.5 (Tikanga) when running the Hg benchmark.


However, I found that these errors disappear when VASP is compiled with no optimization "-O0".

After adding nonlr.o nonl.o to the "OBJ_HIGH =" line , changing the line "OFLAG_HIGH = $(OFLAG)" to "OFLAG_HIGH = -O3" and the line "OFLAG=-O3" to "OFLAG=-O0" in my makefile. I get a vasp gamma point version which does not generate the ZHEGV warnings and has acceptable performance.

It would be nice if the option to use the ZHEEVX routine rather than ZHEGV routine were available in rmm-diss.f like in the subroutines davidson.F, subrot.F and wavpre_noio.F. This would probably be a good fix.

[fish]

The routine wave_high causes the warning messages in EDDRMM when compiled with -O3 in the parallel gamma point version of VASP 5.2.8.

Adding the text "wave_high.o" after the line containing "OBJ_NOOPT=" in the makefile turns the error messages off with little impact on performance.