Installation on IBM Pseries 5, 64 bit

[saurabhroy]

Dear All,
Has anyone tried VASP on an IBM Pseries 5, 64 bit, the compiler I have used is a IBM Advanced XL fortran (ver. 9.1). There seems to be some problem in compilation (syntax errors in the files).

[Jukkis]

Hi all,

I will put my story with the IBM pSeries 690 + VASP here, too.

I compiled the VASP 4.6 as serial version with xlf90 compiler (version 8.1.1.4, as reported by lslpp -L) using the standard makefile.rs6000. FFT3D = fft3dfurth.o fft3dlib.o was selected from the three available options and the code seems to work fine with the test calculation in the benchmark.tar.gz file.

However, the efforts making VASP run parallel using MPI on the same computer have been unsuccessful so far. I started from scratch (different directory, untarred the files again) using the makefile.rs6000 as makefile. When I change the compiler line to mpxlf90_r and specify -DMPI on CPP line of the makefile, I get into trouble. It seems that I am unable to make any of the choices for FFT3D work.

The first attempt with FFT3D=fft3dessl.o ftflop.o
fails and the compiler is complaining about missing symbols (MAPSET, FFTMAKEPLAN, FFTCHK_FURTH,CFTTAB,CFFTML)

The second attempt with (I changed the F77 line to mpxlf77_r) FFT3D=fft3dfurth_f77.o fft3dlib_f77.o
fails with complaints about missing symbols (MAPSET,FFTMAKEPLAN)

The third attempt with FFT3D=fft3dfurth.o fft3dlib.o
fails with error about missing symbol (MAPSET)

Any ideas how to proceed would be greatly appreciated.


Jukka-Pekka Jalkanen

[mstephan]

Hi,

it is better to use makefile.sp2 as an basis for IBM systems like the pSeries. Here you find the correct parameter settings for a parallel version.

For FFT you have to choose

Code: Select all

FFT3D   = fftmpi.o fftmpi_map.o fft3dlib.o

also you have to add a "-lmpi" to the "LIB = " statement.


Michael Stephan

[j.meyer]

Michael Stephan is right:
FFT3D = fftmpi.o fftmpi_map.o fft3dlib.o
works for me too for the parallel version of VASP
on an IBM pSeries 690 machine. It should suffice
to replace the corresponding line in makefile.rs6000
in combination with the other modifications mentioned
to compile the parallel version.

Using
FFT3D = fftmpi_map.o fft3dfurth.o fft3dlib.o
should also link without problems. I have not tried
that so far. Judging from the comments in
fft3dessl+furth.F of current VASP (vasp.4.6.28 25Jul05)
concerning the scaling behaviour of ESSL and fft3dfurth
it should be even better if one could link against this
library. But since VASP 4.6 this seems to be no longer
possible. Linking with
FFT3D = fftmpi_map.o fft3dessl+furth.o fft3dlib.o
fails because the subroutine fftmakeplan is missing
(by the way it is also missing in fft3dessl .F).
Maybe it has just been forgotten to add those six
lines to fft3dessl .F and fft3dessl+furth.F when
this subroutine was introduced in VASP 4.6.
(The name reminds me of the concept that is used
in FFTW...)

However, the question is whether using these FFT
implementations would result in better performance
on the IBM machine. It would be interesting if
someone has experiences with their performance
in previous VASP versions. On the other hand maybe
the developers deliberately did not update fft3dessl .F
and fft3dessl+furth.F because they are sure that this
"fftmakeplan concept" (in combination with the standard
FFT implementations) outperforms anything else.

Jörg Meyer

[mqcai]

Dear :
I am a new vasp user. I have installed the vasp code in the IBM p690 successfully with the parallel edition.

Now I want to run the example H2O. how to run paraell calculation with vasp in IBM p690?

ERROR: 0031-808 Hostfile or pool must be used to request nodes.

Thank you in advance!

[admin]

this error is not vasp-related. Please check how to write an appropriate job-runscript for your parallel environment