Wrt time to solution how much I/O does your job do? Is the filesystem only accessible by you, for instance local disk on the nodes, or shared by all users? — Ian Bush Sep 15 at 17:23

Wrt to different energies if you use a given number of MPI processes and threads do you get the same energy, or does the energy vary between different runs using exactly the same set up? — Ian Bush Sep 15 at 17:25

@IanBush Thanks for taking the time out and sorry as I am a bit new to using HPC systems I might say things that could be wrong. Not entirely sure what you mean by I/O but if you mean iterations then 66,65,66,64(in the same order as the table) so the number of iterations is not increasing but the time of each iteration is increasing. I think that it is shared among users. I can see multiple folders with different names when I do change directory out of my root folder. So I think the filing system is shared with multiple useres — Parmeet Singh EP 066 Sep 15 at 18:18

@IanBush I am not explicitly choosing the number of threads or the number of prcesses. But I the number of cores remains same throughout all the calculations. I have also attached the script I was using to submit the jobs. Perhaps that will answer your questions better. The same script has been used for all the runs — Parmeet Singh EP 066 Sep 15 at 18:21

How large is your system? because you set LREAL = Auto. Your fourth run gave very different total energy, so you better check it carefully. I am not sure whether it is because perovskite system has complex potential energy surface that makes it difficult to find the minimum. How about not setting the IALGO tag? Better set EDIFFG = -0.01 if you do not have a very big system (say ~1000 atoms). — Memories Sep 15 at 21:52

@Memories I currently am trying to run with mBJ-LDA functional a big perovskite super cell with about 80 atoms a lot of which are heavy metals. So the reason I am trying to optimize my calculations is to just make this more efficient. Considering this I think LREAL = FALSE will be a bit too much and I can not set EDIFFG = -0.01 . I am actually a bit unclear about this but what does not setting the IALGO tag accomplish wouldn't I have to do another calculation with IALGO mentioned using the converged WAVECAR. — Parmeet Singh EP 066 2 days ago

OK, I did not notice you are not relaxing the geometry. Nevertheless, if there is no memory bottleneck, I would not agree your calculations are heavy, since they can finish in less than an hour. — Memories 2 days ago

@Memories I agree these calculations are not heavy at all. I am currently using PBE in order to optimize the parallelization, the inputs(NPAR and KPAR) which take the least time will be directly used for mBJ-LDA calculation. The first part(Doing this in multiple steps) for mBJ took Total CPU time used (sec): 68464.703, User time (sec): 48166.752, System time (sec): 20297.948, Elapsed time (sec): 631307.212. And this is just the first step and I have to do the same calculation for more than 10 compounds — Parmeet Singh EP 066 2 days ago