Rough estimate about speed of Casino

General discussion of the Cambridge quantum Monte Carlo code CASINO; how to install and setup; how to use it; what it does; applications.
Post Reply
amartyabanerjee
Posts: 3
Joined: Tue Nov 19, 2013 12:33 am

Rough estimate about speed of Casino

Post by amartyabanerjee »

Hello,
I have been looking into various quantum chemistry methods and I wanted to ask a quick question about Casino. Given that it is a very well optimized and highly scalable code, I was wondering: How long can I expect a 3D periodic crystal, containing about 50 electrons per unit cell to take, on say 1000 or so cpu cores? The typical elements in the unit cell I have in mind are iron and/or copper and I would like to achieve about 0.01 ev/atom certainty/accuracy in the total energies. A rough estimate on the total cpu-hours I can expect such a calculation to take using Casino would be fine.

Thanks in advance.

- Amartya
Neil Drummond
Posts: 113
Joined: Fri May 31, 2013 10:42 am
Location: Lancaster
Contact:

Re: Rough estimate about speed of Casino

Post by Neil Drummond »

Dear Amartya,

Depends on how exactly you perform your calculations. Let's assume you use a supercell containing about 400 electrons, a Slater-Jastrow wave function, a time step of 0.01 a.u. and 4 configurations per core, with 1000 cores, in your DMC calculations. You will probably need to carry out several thousand time steps (dominated by equilibration time). Let's say 5000 steps. This will probably take two or three hours of walltime on your ~1000 cores.

You will probably want to use two different time steps to extrapolate to zero time step. You may also want to twist average (average over offsets to the grid of k vectors) at proportionate cost to the number of offsets.

In addition you will need to optimise a wave function, which will require much less CPU time but somewhat more human time.

Hope this helps,

Neil.
amartyabanerjee
Posts: 3
Joined: Tue Nov 19, 2013 12:33 am

Re: Rough estimate about speed of Casino

Post by amartyabanerjee »

Hi Neil,

Many thanks for your reply. Just to be sure, for the total time estimate you suggested (2000 - 3000 cpu hours) you really meant 400 electrons in the unit cell and not 50 (which I had asked originally) ? I guess I am quite surprised by the fact that it seems to be quite fast... I am aware of routine LDA-DFT calculations on non-periodic cluster systems containing ~ 500 electrons (of Aluminum say, using bulk-fitted local pseudopotentials) taking about 30 cpu hours or so to reach about 1-2 mev/atom convergence.

Please let me know. Thanks again !

- Amartya
Neil Drummond
Posts: 113
Joined: Fri May 31, 2013 10:42 am
Location: Lancaster
Contact:

Re: Rough estimate about speed of Casino

Post by Neil Drummond »

Dear Amartya,

If you are interested in an energy per atom then the CPU time for DMC statistics accumulation to achieve a given error bar is linear in the number of particles in the supercell. If you are interested in total energies, e.g., because you want to calculate points on a band structure, then the time taken grows as the cube of the number of particles. Actually, for the system size, target accuracy and number of cores you suggest, I expect that equilibration will dominate your CPU time.

The 400 electrons came from an assumption that you would use a 2x2x2 supercell.

As I mentioned, the details of your treatment of finite-size effects will affect the timing estimates. In particular, for a metallic system you may want to twist-average (average over offsets to the grid of k vectors). If your CPU time is dominated by equilibration, the cost will be proportional to the number of offsets considered. (However, there are steps that one can take to reduce the cost of equilibration.)

As an example, I have recently calculated the energy per atom of 96-atom cells of various phases of solid hydrogen. To achieve a (very small) error bar of 0.00009 eV / atom in a DMC run at time-step 0.01 a.u. with a Slater-Jastrow trial wave function using the bare Coulomb potential for the hydrogen atoms required 12,288 core hours.

Best wishes,

Neil.
amartyabanerjee
Posts: 3
Joined: Tue Nov 19, 2013 12:33 am

Re: Rough estimate about speed of Casino

Post by amartyabanerjee »

Thanks for the detailed explanation Neil !
Post Reply