PC hardware specifications for complex simulations

Dear FLUKA experts,
During my work, I need to simulate complex geometries for shilding and radiation protection calculations. Those simulations tend to take up to several weeks long for buildings with three floors, for example, in order to aquire adequate statistics.

From you experience, what should be taken into account regarding PC hardware specifications, consedering my main interest is reducing the calculation time?
I am aware of the biasing option in FLUKA, but it is not always possible to implement it in my case.

Currently I have a machine with ~200 GB RAM and 44 cores, with an SSD disk.

Thank you in advance,
Hen Shukrun

Dear Rachel,

the speed of the FLUKA simulations mainly depend on the speed and number of CPU cores. Usually the complex simulations are run on HPC clusters, where one can access hundreds or even thousands of CPUs at a time.

If you don’t have access to a cluster, you can still try to optimize your geometry. For example:

  • Do you need the whole building simulated?
  • Are you getting good statistics everywhere, or only at the area of interest?

I hope this helps.


Pay also attention on L1,2,3 cache sizes. Your memory must be fast: fast cpu with slow memory runs slow. Due to random nature of Monte Carlo most references to memory give cache misses therefore memory performance really matters.
Be also careful with systems where cores split between two sockets due to the NUMA issues.

several weeks long

Wow! Use variance reduction. In the analogue calculations the error is inversely proportional to sqrt(number of incident particles) which means that at first it drops fast but then only slowly decreases with number of incident particles making the rest of your calculation highly inefficient. Correct use of variance reduction can give you orders of magnitude increase in speed.

I am aware of the biasing option in FLUKA, but it is not always possible to implement it in my case.

I can’t easily imagine a situation where biasing would not help, but I’ve seen many problems where the absence of variance reduction would cause undersampling of the phase space and therefore bias your estimator.

If you can’t solve your problem within several hours you are doing something wrong.