My Community

Hi,

I am running an MD simulation with large number of atoms and would like to limit how much VASP writes as some time is spent during IO. I know NWRITE and ML_OUTPUT_MODE, but I wonder whether there is a way to go further. Something like only output XDATCAR or hd5 file nothing more. If possible, this can make each MD step faster?

Regrds,
Burak

Hello Burak,

another INCAR tag to consider is ML_OUTBLOCK. With this tag you can limit the output of atomic positions to every nth step, e.g., will result in an XDATCAR which will contain the positions only for every 10th step. Together with the other tags you mentioned you should be able to further decrease the disk output. Unfortunately, there is not yet a way to disable some remaining lines which are written every time step. For example, the ML_LOGFILE and the OUTCAR file will still contain sections which are written every step (e.g. temperature, total energy,...). We intend to improve this situation in upcoming releases. However, with the use of ML_OUTBLOCK there should be no more output which writes per-atom information to files in each step (tested with VASP.6.4.2). Whether this will really show in your timings I cannot tell because it will depend on the system size, computational demand of the machine-learned force field, hardware, etc.

I have to admit that the documentation of the tags ML_OUTBLOCK, ML_OUTPUT_MODE and NBLOCK and their interplay is somewhat confusing. I will try to update the Wiki in the coming days, sorry for the inconvenience.

All the best,
Andreas Singraber

Dear Andreas,

Thanks for the answer. I need the XDATCAR at every time step so ML_OUTBLOCK is not an option, but I confirm that it works and reduces total time significantly. In this case it seems I need to wait for the next version.

Regards,

Dear Burak,

well, but then the amount of data that is written to disk will usually be dominated by the XDATCAR file anyway because this scales with the number of atoms while other files just write a few lines (e.g. for potential energy,...) per step. One possibility for the future may be to block the disk I/O, i.e., to keep past atom positions in memory until we write them to XDATCAR every nth step. This I have to discuss with my colleagues first, I am not so sure if this will bring a significant speedup.

Best,
Andreas Singraber