NSW is the number of ionic steps. There really isn't anyway to know a priori how many ionic steps will be needed to converge the forces/stresses on the system. I usually try to choose NSW to be something that will ensure my run will fit in the wall clock of the machine I'm using. However, unless the job says:

"reached required accuracy - stopping structural energy minimisation"

Then you'll need to copy the CONTCAR to the POSCAR and continue the relaxation (it is also helpful to have VASP read in the CHGCAR/WAVECAR from the previous run).

The bigger issue that you seem to be having, which may be leading the "random energies" you claim is that the calculation isn't converging within each electronic minimization at each ionic step. There are several things that can contribute to this. For example, your starting geometry may be unreasonable. I also suggest you take a look at the following entry of the VASP manual relating to the mixing tags:

http://cms.mpi.univie.ac.at/vasp/vasp/Mixing_tags_IMIX_INIMIX_MAXMIX_AMIX_BMIX_AMIX_MAG_BMIX_MAG_AMIN_MIXPRE_WC.html

Sometimes, the defaults need to be adjusted in order for the calculation to converge properly:

"you must decrease BMIX if the mean eigenvalue is larger than one, and increase BMIX if the mean eigenvalue is smaller than one"

Here you can find the mean eigenvalue by searching in your OUTCAR for gamma_mean.

Hope this helps.

Thank You Dr. Andrew O'Hara for nice discussion.

It is very useful.

Dr. Andrew O'Hara

It is the right mean eigenval which you are talking. I searched from OUTCAR.

eigenvalues of (default mixing * dielectric matrix)

average eigenvalue GAMMA= 1.0695

Yes, this is the eigenvalue. You'll want to try decreasing the bmix parameter a little and it may help assuming the geometry is fine.

Dear Dr. Andrew I want to ask one thing after reached required accuracy - stopping structural energy minimisation" Then we need to copy the CONTCAR to the POSCAR.

It is the optimized structure and we can do further any required calculation or we have to change volumes in INCAR and taking corresponding energies and by plotting E-V curve we have to see at which volume it is showing minimum energy that is the optimized structure.

Because when I am doing structural optimization of my system it is showing reached required accuracy - stopping structural energy minimisation" at different volume and when I am taking energies at different volumes then minimum energy is coming at different volume and corresponding CONTCAR showing different lattice constant.

So I want to know what is the best way to optimizing structure by single run taking proper NSW value upto reached required accuracy - stopping structural energy minimisation" or the same thing for different volumes and plotting E-V curve and at which volume it is showing minimum energy that is the optimized structure.

In general, you'll find that the energy is slightly different between these two approaches. Depending upon your ENCUT/kpts, it may be inconsequentially small, for example if you've chosen the ENCUT and k-point grid such that the total energy is less than 1 meV / atom or molecular unit and the difference you find is below 1 meV it doesn't really matter.

The primary reason for this discrepancy is described by Pulay Stress. Essentially, this is a force due to the incompleteness of the basis set used. If you intend to use the default ENCUTs of the pseudopotentials, then you'd want to increase this number by ~30% for relaxations (of course if you have already chosen significantly higher ENCUTs for convergence then this isn't as big of an issue). A little bit more can be found here in the manual:

http://cms.mpi.univie.ac.at/vasp/vasp/Volume_vs_energy_volume_relaxations_Pulay_Stress.html

That being said, in regards to your follow up question of which method is preferred? For me it depends upon the system I'm working with. Almost always for cubic systems I will do the energy verse volume curve by hand and sometimes for tetragonal cells as well (with internal coordinates relaxed of course). However, anything beyond this begins to become too cumbersome due to the increased degrees of freedom to do by hand and I will generally relax the lattice vectors with ISIF = 3 making sure my ENCUT and k-points (k-points especially for metals) are well converged.