Initialisation of the calculation
Posted: Thu Jun 09, 2016 2:29 pm
Dear all,
I am trying to benchmark my system to find out the most efficient parallelisation scheme. I have used a 64-atomic SQS supercell to mimic my most typical calculations. In order to run the identical calculations, all were initialised with ISTART=0, INIWAV=1 (to get rid of any randomness) and ICHARG=2. I was measuring the time of only 1 electronic step (NELM=1).
When I analysed the resulting energies, I however found that there were large (>10%) differences between individual settings of NPAR, KPAR, OMP_NUM_THREADS, and -np in mpirun. I am puzzled by this, as I was expecting the calculation to run a deterministic fashion for given initial conditions, which, in my understanding should be the same for all calculations (jellium model for \psi and superposition of atomic charge densities for \rho).
Since the energies differ so much I do not this the above described procedure makes much sense and cannot be used for benchmarking. I, however, have no clue why do the energies differ so much. My further analysis revealed that this difference comes from different eigenvalues in each specific case, but why is this happening for the identical initial conditions?
Many thanks in advance for any explanation/hint.
David
I am trying to benchmark my system to find out the most efficient parallelisation scheme. I have used a 64-atomic SQS supercell to mimic my most typical calculations. In order to run the identical calculations, all were initialised with ISTART=0, INIWAV=1 (to get rid of any randomness) and ICHARG=2. I was measuring the time of only 1 electronic step (NELM=1).
When I analysed the resulting energies, I however found that there were large (>10%) differences between individual settings of NPAR, KPAR, OMP_NUM_THREADS, and -np in mpirun. I am puzzled by this, as I was expecting the calculation to run a deterministic fashion for given initial conditions, which, in my understanding should be the same for all calculations (jellium model for \psi and superposition of atomic charge densities for \rho).
Since the energies differ so much I do not this the above described procedure makes much sense and cannot be used for benchmarking. I, however, have no clue why do the energies differ so much. My further analysis revealed that this difference comes from different eigenvalues in each specific case, but why is this happening for the identical initial conditions?
Many thanks in advance for any explanation/hint.
David