In addition to what Micheal said, there is also the nature of your system which will play an important role in the entire convergence story. Since the k-points are used in the integration scheme that calculates energy of a system, the actual bandstructure of the system will also play an important role in how large the k-point set will need to be to reach the same level of convergence for systems of the same size.

This is why metals for example require a relatively larger k-point set than semiconductors or insulators. But also systems which are known for their peculiar bandstructure  such as graphene will require a denser or even a specialized grid to accurately sample the brillouin zone. In the end, there is not really a simple one-shot formula to give you a measure of accuracy based solely on parameters provided.

Effectively you raise the cutoff energy and k-mesh density until you reach convergence. You then use that as reference point, at a certain point the changes will have decreased enough to estimate an upper boundary on the mistake (hence the ~). 

This behavior will vary depending on the property you perform this calibration test on. Energy difference will for instance converge much faster than absolute energies, due to error cancellation.

In function of the cutoff energy the absolute energy will decrease. The required value is typically linked to the specific elements present (and potentials used), not as much of the geometry). In VASP for instance there is an ENMAX present in the POTCAR, but this is typically an underestimation for the accuracies expected nowadays. Typically for PAW potentials 400 or 500 eV is good for normal potentials and 850 eV for hard ones (_h). Higher than 850 eV can sometimes cause warnings (before the first electronic step), dependent on the exact elements used. 

K-mesh of course depends on the size of the cell. At low density it will first oscillate then it will flat line and not change too much. When using a gamma-centered grid it's best to just check the uneven numbers (3x3x3, 5x5x5, 7x7x7, etc) to ensure it is symmetrical around the gamma point. Keep in mind that many k-points are symmetry-equivalent, adding more of those will not change the energy. The number of inequivalent k-points can be found in VASP in the IBZKPT file. It is really convergence of energies (or energy differences) with respect to this number that matters.

For supercell size, given sufficient k-points and encut (k-points can be divided by the amount of unit cells in the supercell), you once again raise it to larger cells and study how the energy differences you are interested in evolve.

Always keep in mind that not all properties converge equally fast, though many are similar. Optical properties and frequencies and such require significantly higher settings. In those cases also keep in mind things like FFT grid sizes and other parameters that may change the energy (Prec=Accurate, LREAL=.FALSE., etc in VASP)

In addition to what Micheal said, there is also the nature of your system which will play an important role in the entire convergence story. Since the k-points are used in the integration scheme that calculates energy of a system, the actual bandstructure of the system will also play an important role in how large the k-point set will need to be to reach the same level of convergence for systems of the same size.

This is why metals for example require a relatively larger k-point set than semiconductors or insulators. But also systems which are known for their peculiar bandstructure  such as graphene will require a denser or even a specialized grid to accurately sample the brillouin zone. In the end, there is not really a simple one-shot formula to give you a measure of accuracy based solely on parameters provided.

Thank you very much for your detailed answers.

I have another question about the Kpoints mesh.

Taking SrTiO3 structure as example, the atomic position is:

Sr 0.5 0.5 0.5

Ti 0 0 0

O 0.5 0 0

O 0 0.5 0

O 0 0 0.5

I used 6X6X6 K-mesh for its unit cell. And its high symmetric Kpoints path are along with the  body center, face center, mid of edge and vertex.

When I added it to a supercells (sqrt(2) X sqrt(2) X 2, total 20 atoms, lattice parameters a' = sqrt(2)*a, c' = 2*c), I used 6X6X3 Kmesh.

Sr        0.5         0         0.3535

Sr        0            0.5      0.3535

Sr        0.5         0         1.0605

Sr        0            0.5      1.0605

Ti        0            0          0

Ti        0.5         0.5       0.707

Ti        0.5         0.5       0

Ti        0            0          0.707

O        0.25       0.25      0

O        0.75       0.25      0

O        0.25       0.75      0

O        0.75       0.75      0

O        0            0           0.3535

O        0.5         0.5         0.3535

O        0.25       0.25       0.707

O        0.75       0.25       0.707

O        0.25       0.75       0.707

O        0.75       0.75       0.707

O        0            0           1.0605

O        0.5         0.5        1.0605

My question is in this case how to choose the high symmetrical Kpoint path to plot the band structure?

The symmetry of cells can be divided into space groups, forr each space group there is a standardized pathway known. There are various placed where these are documented. The easiest way is to use aconvasp (either the tool or website), it gives you the KPOINTS file containing the path for a given POSCAR (and a picture). 

In steps:

1) go to http://materials.duke.edu/awrapper.html

2) click Kpath under symmetry

3) paste in your POSCAR and hit start

For general data about symmetry, wyckoff positions, space groups and Brillouin zones I recommend http://www.cryst.ehu.es/ .

PS: XCrySDen also offers a graphical way to build k-paths (by selecting points in the Brillouin zone, but it does not natively support VASP format.

In addition to previous statements on the k-point issue:

To carry out a supercell calculation with same accuracy as your unit cell. The number of k-points should be scaled, otherwise one calculation will be more accurate than the other.

Also, except you are introducing dopants or carrying out some modifications to your conventional unit cell, it is advisable to carry out your band structure calculation on your unit. You would have the benefits of faster calculation and more accurate band structure because you can increase the number of your k-points and other convergence criteria.

Michael has provided a good answer on how to determine the kpoints for both unit cell and supercell calcations