I wanted to know how do you determine how many cores and noes you will need to perform an DFT analysis for particular molecules??
I normally use two cores for all my calculations and it pretty much serves the purpose. But I was curious if there is any ratio for that??
#DFT #VASP #HPC
There's not really a way to determine this a priori, because it is dependent on the HPC.
For isolated molecules, generally speaking, your code should run reasonably fast without parallelization. However, for surfaces, this is a different.
1) For one thing, with periodic DFT, we normally simulate a supercell of a bulk structure cleaved along a plane. If you have a four-layered, 4 x 4 fcc (111) surface for example, you will have 64 atoms total.
2) Often, the constituent atoms of surfaces have a greater number of valence electrons per atom than atoms which make up typical molecules.
For example, suppose I want to calculate the adsorption of benzene on a four-layered, 4 x 4 Pt (111) surface. The benzene molecule has a total of 30 valence electrons (6 x 4 + 6 x 1). Pt, however, has 10 valence electrons according to most DFT setups (the number of valence electrons is dependent on the pseudopotential you're using -- different pseudopotentials will sometimes consider differing number of electrons to be "core" vs. "valence"). Thus, the platinum surface in this case would have a total of 640 valence electrons.
Computation time of DFT generally scales as O(Ne3). For a brief description of this, check out this answer on StackExchange. https://scicomp.stackexchange.com/questions/5515/how-does-density-functional-theory-scale-with-system-size
For calculations involving surfaces, as a general rule of thumb, you will probably need to run in parallel on an entire node of the HPC (e.g., ~20 cores) to get convergence in less than a day of run time. This is very approximate and will be highly system-dependent.
There's not really a way to determine this a priori, because it is dependent on the HPC.
For isolated molecules, generally speaking, your code should run reasonably fast without parallelization. However, for surfaces, this is a different.
1) For one thing, with periodic DFT, we normally simulate a supercell of a bulk structure cleaved along a plane. If you have a four-layered, 4 x 4 fcc (111) surface for example, you will have 64 atoms total.
2) Often, the constituent atoms of surfaces have a greater number of valence electrons per atom than atoms which make up typical molecules.
For example, suppose I want to calculate the adsorption of benzene on a four-layered, 4 x 4 Pt (111) surface. The benzene molecule has a total of 30 valence electrons (6 x 4 + 6 x 1). Pt, however, has 10 valence electrons according to most DFT setups (the number of valence electrons is dependent on the pseudopotential you're using -- different pseudopotentials will sometimes consider differing number of electrons to be "core" vs. "valence"). Thus, the platinum surface in this case would have a total of 640 valence electrons.
Computation time of DFT generally scales as O(Ne3). For a brief description of this, check out this answer on StackExchange. https://scicomp.stackexchange.com/questions/5515/how-does-density-functional-theory-scale-with-system-size
For calculations involving surfaces, as a general rule of thumb, you will probably need to run in parallel on an entire node of the HPC (e.g., ~20 cores) to get convergence in less than a day of run time. This is very approximate and will be highly system-dependent.
It strongly depends on the size of the molecule. In general VASP parallelizes at different levels. It parallelizes at the level of plane waves, bands and k-points. Now for a molecule the last point is irrelevant as you only have the Gamma-point anyhow (so forget about KPAR). Inside a node, if you do not set NPAR, VASP will parallelize via the underlying math-library (mkl or lapack) to take up any cores available. This can actually kill your calculation if you have too few electronic bands (or electrons, try to have more than 5 bands/core to keep safe).
In case you have a somewhat sizeable molecule (say 150 atoms and 500 electrons), running on a machine with 20 cores, you may want to consider using multiple nodes. Here you start using the NPAR (or NCORE) tag, setting it equal to the number of nodes used.
Yes, there is a calculation for that.
From the pseudopotential file (POTCAR for VASP, and *.UPF for QE), count the number of valence electrons and the number of bands. This can also be found from a simple SCF calculation. Look at the tag NBANDS in VASP.
Ideally the number of cores to be used = NBANDS/8
Therefore, if NBANDS=300, then # of cores = 300/8 = 37.5 ~ 38. So, allot 38 cores in 2 nodes. You can do this by (i) allotting 19 cores in each node (deprecated) or (ii) 32 cores per node which equals 64 cores total (will be redundant). So, stick to the first method.
If you need the results very fast then # of cores = NBANDS/[4 OR 6]. Don’t go less than 4 as it will burn out the resources for no reason.
**Some jobs require very large memory due to heavy atoms and tight criteria. Then, decrease the number of cores and increase the number of nodes. This then invokes the node memory as not only of the local cores.
For example, for any job if you need 192 cores, then DON’T use 192/32= 6 nodes. Rather use, 192/16= 12 nodes. BUT KEEP IN MIND THAT WILL BURN A LOT OF MEMORY AND USE IT ONLY IF ABSOLUTELY NECESSARY! **
Try to avoid KPAR tag unless needed to speed up calculations. KPAR eats up too much of core hours too.
Set NPAR= # of nodes. This is quite optimum.
- Badges
- Science topic
- Similar topics
- Biological Science
- Kinesiology
- Running