how to set the parallel parameters (paral_kgb,npband etc)?
Posted: Sat Aug 20, 2011 5:06 pm
hi,everyone!
i have read all of the topics of parallel setting in this forum,but i am still confused.
for gs calculation, abinit can parallel on three levels which are k-point, G-vector and band.
The corresponding parameters are npkpt, npfft and npband, respectively.
in order to understand how to sett parameters, i have done several calculations.
Experiment 1:
A small system , with only one atom (Fe) in the primitive cell. in fact, it's the example of spin in abinit official tutorial.
i changed the k-point setting, and used a bigger one (14*14*14).
first, i did the calculation with paral_kgb=0,which is default setting by abinit. The Overwall time is 315.2 sec.
second, i opened parallel on kgb.(npkpt=npfft=1.npband=5). The Overwall time is 644.6 sec.
abinit version is 6.8.1, the computer is small dell workstation with 2 xeon( 8 processors). I had just used 5 processors in experiment 1.
Experiment 2:
a bigger system with 17 atoms( 16 fe and one N),and k-point setting is 2*2*2.
first, i did the calculation with paral_kgb=0,which is default setting by abinit. The Overwall time is 31122 sec.
second, i opened parallel on kgb.(npkpt=npfft=1.npband=8). The Overwall time is 13424 sec.
abinit version is also 6.8.1 and the computer is the same with Exp.1. This time ,used 8 processors.
compare Exp. 1 with 2, i get the result : for system with larger k-point grid , it is better to parallel on k-point (which is default of abinit). for system with a lot of atoms, the band is the main factor, it is better to parallel on band.
i also have done several calculations in order to understand npfft, but got no idea about it until now .
I want to ask following questions.
a. Do i have got the correct result about npkpt and npband setting?
b. Can someone tell me how to tune npfft?
c. For a system, we know the atoms number, the band number and k-point grid., how to tune npkpt, npband and npfft to get the optimal setting?
think you for your time!
hhwj
i have read all of the topics of parallel setting in this forum,but i am still confused.
for gs calculation, abinit can parallel on three levels which are k-point, G-vector and band.
The corresponding parameters are npkpt, npfft and npband, respectively.
in order to understand how to sett parameters, i have done several calculations.
Experiment 1:
A small system , with only one atom (Fe) in the primitive cell. in fact, it's the example of spin in abinit official tutorial.
i changed the k-point setting, and used a bigger one (14*14*14).
first, i did the calculation with paral_kgb=0,which is default setting by abinit. The Overwall time is 315.2 sec.
second, i opened parallel on kgb.(npkpt=npfft=1.npband=5). The Overwall time is 644.6 sec.
abinit version is 6.8.1, the computer is small dell workstation with 2 xeon( 8 processors). I had just used 5 processors in experiment 1.
Experiment 2:
a bigger system with 17 atoms( 16 fe and one N),and k-point setting is 2*2*2.
first, i did the calculation with paral_kgb=0,which is default setting by abinit. The Overwall time is 31122 sec.
second, i opened parallel on kgb.(npkpt=npfft=1.npband=8). The Overwall time is 13424 sec.
abinit version is also 6.8.1 and the computer is the same with Exp.1. This time ,used 8 processors.
compare Exp. 1 with 2, i get the result : for system with larger k-point grid , it is better to parallel on k-point (which is default of abinit). for system with a lot of atoms, the band is the main factor, it is better to parallel on band.
i also have done several calculations in order to understand npfft, but got no idea about it until now .
I want to ask following questions.
a. Do i have got the correct result about npkpt and npband setting?
b. Can someone tell me how to tune npfft?
c. For a system, we know the atoms number, the band number and k-point grid., how to tune npkpt, npband and npfft to get the optimal setting?
think you for your time!
hhwj