unusual slow calculation with larger systems

Questions regarding the compilation of VASP on various platforms: hardware, compilers and libraries, etc.

Moderators: Global Moderator, Moderator

Post Reply
Message
Author
frontleg

unusual slow calculation with larger systems

#1 Post by frontleg » Tue Jun 04, 2013 10:36 pm

Hi, All,

We tried to compile the vasp 5.2 and 5.3 on new linux cluster (http://www.rcac.purdue.edu/userinfo/resources/carter/). It worked quite well at the beginning. But after we increase the system size (e.g. from non-spin polirized MgO to spin polirized and +U corrected FeO; from small vacuum to large vacuum). The calculation became extremely slow in some cases. In such slow cases, OUTCAR file showed that the "EDDIAG:" and "ORTHCH:" time is much longer (hundreds seconds and even > 1000s) than it in the normal case (< 10s). The "RMM-DIIS:" time do not have a big different in all the cases.

We checked the memory usage and intra-nodes communications and it turned out no limitations from the hardware.

Could anyone give us some hint on how to solve this problem?

Thank you in advance!

zhao
Last edited by frontleg on Tue Jun 04, 2013 10:36 pm, edited 1 time in total.

admin
Administrator
Administrator
Posts: 2922
Joined: Tue Aug 03, 2004 8:18 am
License Nr.: 458

unusual slow calculation with larger systems

#2 Post by admin » Wed Jun 05, 2013 11:29 am

Expanding the vacuum is not "free of charge." The number of plane waves scales with the cell volume. With increasing volume the PW basis set increases considerably and this leads to the increase of the computer time.
Last edited by admin on Wed Jun 05, 2013 11:29 am, edited 1 time in total.

frontleg

unusual slow calculation with larger systems

#3 Post by frontleg » Wed Jun 05, 2013 3:59 pm

Thank you very much for the reply. I am sorry that I did not explain my problem very clear.

My problem is the computational time is randomly slow when we have a larger system. For example, in one of my calculation, in the first SCF LOOP, the speed is normal, e.g. the time consuming for one electronic step is as following:

POTLOK: cpu time 0.85: real time 0.91
SETDIJ: cpu time 1.73: real time 1.73
EDDIAG: cpu time 9.22: real time 9.27
RMM-DIIS: cpu time 13.58: real time 13.61
ORTHCH: cpu time 4.12: real time 4.13
DOS: cpu time 0.00: real time 0.00
CHARGE: cpu time 0.69: real time 0.69
MIXING: cpu time 0.01: real time 0.01
--------------------------------------------
LOOP: cpu time 30.19: real time 30.38

However, from the second SCF LOOP, the calculation became very slow, here is the time for one electronic step:

POTLOK: cpu time 1.47: real time 1.47
SETDIJ: cpu time 1.72: real time 1.72
EDDIAG: cpu time 380.17: real time 381.03
RMM-DIIS: cpu time 12.35: real time 12.38
ORTHCH: cpu time 46.10: real time 46.21
DOS: cpu time 0.00: real time 0.00
CHARGE: cpu time 40.64: real time 40.73
MIXING: cpu time 0.01: real time 0.01
--------------------------------------------
LOOP: cpu time 482.47: real time 483.63

Note that the RMM-DIIS time is nearly the same in these two cases, but others like EDDIAG, ORTHCH, CHARGE became very slow.

We think it might not be the problem of the hardware. So we are wondering whether it is the problem with the compiling of vasp or other related programs.

Could anyone give us some hint on it?

Thank you in advance!

zhao
Last edited by frontleg on Wed Jun 05, 2013 3:59 pm, edited 1 time in total.

Post Reply