Reblock convergence with large number of walkers

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
Kayahan
Posts: 22
Joined: Fri Jun 07, 2013 5:56 pm

Reblock convergence with large number of walkers

Post by Kayahan »

Dear CASINO Users,

As I read from the manual, number of walkers for statistical dmc calculation should be around or more than 1000. However, if I want to use more walkers, say 20000 for example, I was able to get smaller variances by using less number of steps, like one-tenth steps. Although the variance of energy is 8*10^-4 Ha for nearly 100 electron system at 10000 statistical steps (for 20000 walkers), error of standard mean is still around 30 % of standard mean.

It seems advantageous to use large number of walkers if you want to use large number of processors for a short time, but reblocking error requires a some amount of steps to make accurate prediction of the error(it seems like I need to use more than 10000 steps no matter what number of walkers I use for my system).Since number of walkers should not be larger than number of processors, this brings a limitation on number of processors that can be used. Please correct me if I am wrong until here. That is why, I am wondering if there is a good strategy to decrease correlation length. Performing many VMC steps and printing very few of them can be useful I guess as the first step, but I can't think of anything else. Any idea on minimizing CPU run time while minimizing the wall time could be very useful.

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

Re: Reblock convergence with large number of walkers

Post by Neil Drummond »

Dear Kayahan,

For the DMC statistics accumulation stage, to achieve a given error bar you need to sample a given number of independent configurations. In principle the computational expense is the same, irrespective of whether you use a large population and a small number of steps or a small population and a large number of steps. The only way of getting out of this is to use a larger time step.

As you say, using a small number of steps makes reblocking difficult. Furthermore, the number of walkers must be at least the number of cores (unless you want to play with the OpenMP version of CASINO).

The more serious problem with using a huge configuration population is the cost of equilibrating the configurations. There are some tricks for reducing the expense of equilibration, e.g., when twist averaging you can do most of the equilibration at one twist and then do a short re-equilibration for each twist. You can also use the "preliminary DMC" scheme: see the manual.

Best wishes,

Neil.
Post Reply