Configuration's effect on optimization [SOLVED]

[cpromao]

I have two builds of 8.0.8, one on a local machine and one on a HPC cluster. When performing structural optimization using ionmov = 2, sometimes the local build is able to complete the optimization normally but the cluster build shows worsening SCF convergence with each MD step and then crashes. I've attached my input file for BPO4, which exhibits this problem.

The important configuration options for the local build are as follows:

Code: Select all

  * C compiler      : gnu version 5.4
  * Fortran compiler: gnu version 5.4
  * architecture    : intel xeon (64 bits)

  * debugging       : basic
  * optimizations   : standard

  * OpenMP enabled  : no (collapse: ignored)
  * MPI    enabled  : yes
  * MPI-IO enabled  : yes
  * GPU    enabled  : no (flavor: none)

  * TRIO   flavor = none
  * TIMER  flavor = abinit (libs: ignored)
  * LINALG flavor = custom (libs: user-defined)
  * ALGO   flavor = none (libs: ignored)
  * FFT    flavor = none (libs: ignored)
  * MATH   flavor = none (libs: ignored)
  * DFT    flavor = none

while those for the cluster build are:

Code: Select all

  * C compiler      : intel version 16.0
  * Fortran compiler: intel version 16.0
  * architecture    : intel xeon (64 bits)

  * debugging       : basic
  * optimizations   : standard

  * OpenMP enabled  : no (collapse: ignored)
  * MPI    enabled  : yes
  * MPI-IO enabled  : yes
  * GPU    enabled  : no (flavor: none)

  * TRIO   flavor = netcdf
  * TIMER  flavor = abinit (libs: ignored)
  * LINALG flavor = mkl (libs: user-defined)
  * ALGO   flavor = none (libs: ignored)
  * FFT    flavor = fftw3 (libs: user-defined)
  * MATH   flavor = gsl (libs: user-defined)
  * DFT    flavor = libxc+atompaw

If anyone has an idea as to which aspect of the build could be causing this problem it would be greatly appreciated.

[Jordan]

Hi,

Depending on the CPU you have on your localcomputer and on you cluster, this could be due to a vectorization issue with intel16.
Could you try to build abinit with the --enable-avx-safe-mode on your cluster and try again.

Also could you check if the energies and forces are the same the moment the simulations diverge. Or at least identify what physical parameter starts diverging.

Cheers

[cpromao]

Hi Jordan,

Using --enable-avx-safe-mode seems to have solved the problem - thanks!

There is a small difference between energies and forces from the start of the calculation; this persists for 6 MD steps at which point the forces begin to diverge.