ISU / DEC 10TH
Joaquin Peralta, Rupa Dumpala, and Scott Broderick
First simulations / QE• Quantum espresso
• Convergence parameters
• Technical Issues• Undefined ‘input_file’ to add initial velocities.• Defining by modifying source code.• Alternatives VASP/QE Coding
InitialSimulations
Using VASP• Available VASP versions : 5.3.2 and 4.X• Pseudo potentials
• PAW – PBE/LDA• Alternative of request all-electron (very small number of atoms)
• Convergence settings• Molecules with considerable vacuum space• Plane Waves must be around 50% more than the used in classical
crystalline structures (RAM memory)• Partial occupancies (smearing). In particular for the case of molecules
and DM, Fermi-smearing or Gaussian-smearing will be used.• Compilation Issues
• Performance• Gamma Point
Si9 Cluster• Original Adri’s group Si cluster of 9 atoms.• Reducing vacuum space size to improve convergence
and plane waves.• A cubic cell of 30A it was the first option.• Silicon first because :
• Performance Input : ISPIN / ENCUT / NGX• Performance Compiling : Gamma Point / Settings• Vacuum Space, Plane waves and Memory requirements
Si9 / Cluster• Trying to fit best parameters for simulation time
• Technical issues with• Cluster platform fails with NPAR major than eight • Compiled version don’t run with a wrong number of NGX, NGY,
and NGZ• Not considerable performance improvement with more aggressive
compilation• 1 Ionic step/minute 1fs each step 5ps = 83 hours• 1 Ionic step/minute 0.25fs 5ps = 13 days!
Step/hour NPAR 2 NPAR 4 NPAR 8 NPAR 16Regular 46.10/45.8* 57.845** 54.342 -Didn’t work
Gamma -- 60.277** 57.83 -Didn’t work
* Vasp 5.2.11** The simulation didn’t finish.
Si9 - O• Incorporation of Oxygen in the system• Using ISMEAR Associated to the temperature of the input
file from Adri’s group. T = 338 K• Oxygen ENCUT increased the time of the RUN.• Technical initial problems / Times
• Using 64 CPU / NPAR 4 Work ~1 step/min.• Using 64 CPU / NPAR 8 Faster, but suddenly simulation stop.• Using 128 and 256 CPU
• NPAR > 8 Simulation faster ~ 1.5 to 2.0 step/min• Principal problem, after a couple of hours the simulation stop for
different problems• Zombie process (Cluster environment problem)• Setup ENCUT / NGX / etc.
Si9 - O
Si9-O
Si9-O• Temperature response
Si-9• Energy
Next steps• Reduce vacuum space to improve simulation times.• Reduce compiling optimization and disk I/O
• Cluster problem(s)• FFT size grid Better accuracy, slower simulation time.
• Based on old NPAR and CPU studies to compare.