Expected numerical results

[jfredin]

I am building VASP on em64t Westmere processors using the make files documented in the forum using Intel 12.1 compiler, Intel 12.1 MKL and SGI MPT. I want to verify that the build is getting the correct numerical results.

Using the tutorials I get slightly different results that the published results:
http://cms.mpi.univie.ac.at/vasp-worksh ... tation.htm

1_1_Oatom:utorial hands_on:
DAV: 6 -0.31407E+00 -0.86398E-03 -0.18767E-03 14 0.397E-0 0.142E-01
DAV: 7 -0.31422E+00 -0.14198E-03 -0.21862E-04 14 0.149E-0 0.480E-02
DAV: 8 -0.31427E+00 -0.55085E-04 -0.26453E-05 14 0.469E-0
1 F= -.31427624E+00 E0= -.16001392E+00 d E =-.308525E+00
writing wavefunctions

My results 1 mpi thread using 2 socket / 6cores/socket Westmere X5690 @ 3.47GHz 12288KB Cache
DAV: 5 -0.313519571108E+00 -0.29791E-02 -0.29748E-02 28 0.911E-01 0.286E-01
DAV: 6 -0.314539504997E+00 -0.10199E-02 -0.17941E-03 14 0.380E-01 0.142E-01
DAV: 7 -0.314637180692E+00 -0.97676E-04 -0.23578E-04 14 0.153E-01
1 F= -.31463718E+00 E0= -.16037486E+00 d E =-.308525E+00
writing wavefunctions

My result using 12 cores on Westmere X5690 @ 3.47GHz 12288KB Cache
DAV: 4 -0.313513344409E+00 -0.94780E-02 -0.94534E-02 36 0.119E+00
DAV: 5 -0.313520307009E+00 -0.69626E-05 -0.69636E-05 24 0.349E-02 0.286E-01
DAV: 6 -0.314540855907E+00 -0.10205E-02 -0.18034E-03 36 0.222E-01 0.142E-01
DAV: 7 -0.314636474293E+00 -0.95618E-04 -0.21581E-04 24 0.912E-02
1 F= -.31463647E+00 E0= -.16037415E+00 d E =-.308525E+00
writing wavefunctions

I get the same kind of differences on the other hands_on problems.

bench.Hg OSZICAR.ref distributed with input on VASP website
RMM: 10 -0.532326396761E+05 -0.16205E-01 -0.32064E-01 694 0.888E-01 0.358E-01
RMM: 11 -0.532326387202E+05 0.95591E-03 -0.49341E-02 690 0.364E-01 0.898E-02
RMM: 12 -0.532326385045E+05 0.21567E-03 -0.65914E-03 692 0.968E-02 0.548E-02
RMM: 13 -0.532326384716E+05 0.32901E-04 -0.15794E-03 599 0.566E-02
1 F= -.53232638E+05 E0= -.53232641E+05 d E =0.638724E-02

My results from 1 core on Westmere X5690 @ 3.47GHz 12288KB Cache
RMM: 10 -0.532327148448E+05 0.13458E-01 -0.69686E-01 695 0.120E+00 0.550E-01
RMM: 11 -0.532327089541E+05 0.58908E-02 -0.18550E-01 693 0.501E-01 0.247E-01
RMM: 12 -0.532327075118E+05 0.14423E-02 -0.34613E-02 691 0.226E-01 0.756E-02
RMM: 13 -0.532327075990E+05 -0.87187E-04 -0.65477E-03 688 0.823E-02
1 F= -.53232708E+05 E0= -.53232710E+05 d E =0.749678E-02

My result on 12 cores of Westmere X5690 @ 3.47GHz 12288KB Cache
RMM: 10 -0.532327148448E+05 0.13458E-01 -0.69686E-01 695 0.120E+00 0.550E-01
RMM: 11 -0.532327089541E+05 0.58908E-02 -0.18550E-01 693 0.501E-01 0.247E-01
RMM: 12 -0.532327075118E+05 0.14423E-02 -0.34613E-02 691 0.226E-01 0.756E-02
RMM: 13 -0.532327075990E+05 -0.87187E-04 -0.65477E-03 688 0.823E-02
1 F= -.53232708E+05 E0= -.53232710E+05 d E =0.749678E-02

I have tried detuning the whole code to -O0 and this does not change the kinds of differences I see in different tests.

My question is what numerical precision do you expect for verifying the output of VASP and what tests do you use to check that you are getting correct results?

Thanks,
Jan Fredin
SGI

[Dr_Nick]

I noticed similar behaviour with both VASP 4.6.38 and 5.2.12. i get exactly the same numbers as you for bench.Hg (X5690, Intel 12 / MKL / ifftw).

When running in parallel you should set the same NPAR value for all calculations to get identical numbers.