recalculate W for EVGW0?

[xiaoming_wang]

Hello,

I was wondering why for EVGW0 calculations (NELMGW>1), the W would be recalculated for each GW step, as evidenced by the time stamp of the output WFULL* files? My input is as follows:

Code: Select all

ENCUT    =   500 eV
PREC     =   N
LREAL    =   FALSE
GGA      =   PE

ISMEAR   =   0
SIGMA    =   0.05
EDIFF    =   1.e-8

NBANDS = 1024
ALGO = EVGW0
NELMGW = 4
ENCUTGW = 250
PRECFOCK = Fast
NOMEGA = 128 
NBANDSGW = 1024

I noticed that for every GW step, the W matrices would be recalculated and printed out, which wastes a lot of time.

Best,
Xiaoming Wang

[ahampel]

Hi Xiaoming,

thank you for contacting us via the official VASP forum.

I had a quick look at can confirm that the WFULL*** files are indeed written in each iteration. In principle I do agree that W is kept fixed during an EVGW0 calculation and I think I can confirm this when checking the profiling -DPROFILING build VASP makes only one call there. If you grep for GWLOOP in OUTCAR do you observe that the first iteration takes the same time the following ones do, or are the GW steps faster after the first iteration?

I have to confirm this with the team and also double check if there is maybe a technical reason why the files are written out in each iteration.

Have a nice weekend,

Alex H.

[xiaoming_wang]

Hi Alex,

Thanks for your reply.
Yes, the first iteration takes the same time as the following ones. I remembered for older version, say 5.4, the following iterations are usually much faster than the first one and only the first iteration printed out the WFULL* matrices, in GW0 calculations.

Great weekend!

Best,
Xiaoming

[ahampel]

Hi Xiaoming,

I cross checked with M. Kaltak who developed and wrote the GW routines. He confirmed that the WFULL*** files are only touched not written in each iteration, and in profiling it is very clear that XI and W are only calculated in the first iteration. Furthermore, he run a test calculation with VASP 5.4.4 and found that the calculation took the same amount of time:

Code: Select all

OUTCAR.6.5.0-mpiifort:                  Total CPU time used (sec):      153.849
OUTCAR.5.4.4-mpiifort:                  Total CPU time used (sec):      151.697

for some arbitrary test job. So from our side everything seems to be fine. If you can provide timings that proves that for you in 5.4.4 the EVGWO is faster please do provide output and input files for both runs and we will look into it. Please make sure to use the same number of bands NBANDS in each calculation.

Best regards,
Alex