« previous next »
Pages: [1]   Go Down

Author Topic: Phonon Parallelization  (Read 2081 times)

0 Members and 1 Guest are viewing this topic.

Offline sayantanu

  • Heavy QuantumATK user
  • ***
  • Posts: 48
  • Country: in
  • Reputation: 0
    • View Profile
Phonon Parallelization
« on: July 14, 2016, 08:30 »
Dear ATK users,

I'm willing to perform a phonon calculation for a system with 16 atoms in the unit cell. Now as I was checking the tutorial (http://quantumwise.com/publications/tutorials/item/836-silicon-phonon-bandstructure), I got a bit confused about the parallelization. According to that tutorial I should run my calculations with 16x3=48 MPI processes, is it so? We have ATK 13.8 with 16-core license. But the calculation is not getting started when using 4 nodes with 4 cores per node.
Logged

Offline Anders Blom

  • QuantumATK Staff
  • Supreme QuantumATK Wizard
  • *****
  • Posts: 5429
  • Country: dk
  • Reputation: 89
    • View Profile
    • QuantumATK at Synopsys
Re: Phonon Parallelization
« Reply #1 on: July 14, 2016, 13:10 »
I would honestly not attempt that calculation with version 13.8 and just 16 cores, it will be very slow. The newer ATK (esp. 2016) will be able to parallelize this job to 100s of cores for reasonable calculation time.

But what do you mean "not getting started"? Error messages?
Logged

Offline sayantanu

  • Heavy QuantumATK user
  • ***
  • Posts: 48
  • Country: in
  • Reputation: 0
    • View Profile
Re: Phonon Parallelization
« Reply #2 on: July 15, 2016, 15:53 »
The error message is regarding the "symmetrize" and "acoustic_sum_rule" tags; these two are not recognized. I guess these options are not available in ATK 13.8.0, because these two were not present in the generated python file, I added them after reading the tutorial.

But is it possible, in principle, to run the phonon calculation for a system with 16 atoms (i.e. 48 displacements) using 16 cores ? The discussion about parallelization in the tutorial suggests number of MPI processes to be equal to the number of displacements.
Logged

Offline Anders Blom

  • QuantumATK Staff
  • Supreme QuantumATK Wizard
  • *****
  • Posts: 5429
  • Country: dk
  • Reputation: 89
    • View Profile
    • QuantumATK at Synopsys
Re: Phonon Parallelization
« Reply #3 on: July 15, 2016, 16:29 »
Those keywords indeed only work in newer version, on the other hand there are good reasons why they were introduced, since they improve the results quite a bit... Yes, for 16 atoms the calculation will easily parallelize to 48 MPI processes. However, I cannot actually recall if the 13.8 version is parallel over displacements... But you should notice that quite fast because if it is, then the log file will be a bit messy - all DFT calculations are going on at the same time and writing output to the same file (improved a lot in the 2016 edition). If not, then it will parallelize over k-points instead and you still get some benefit.

PS: The latest 13.8 edition is 13.8.2, with many bug fixes over 13.8.2.

Logged

Offline sayantanu

  • Heavy QuantumATK user
  • ***
  • Posts: 48
  • Country: in
  • Reputation: 0
    • View Profile
Re: Phonon Parallelization
« Reply #4 on: July 19, 2016, 12:31 »
Ok....got it. Thank you  :)
Logged
Pages: [1]   Go Up
« previous next »