Scalability question

[Helveg]

Can this generator handle 100,000 or a million particles? What do you estimate the runtime would be using supercomputers (these)? (Using some Or what are the O(...) runtimes for the LS or FBA algorithms?

[VasiliBaranov]

Hi Helveg,

the running time varies with the contraction rate (which determines your final density). See Fig. 2a and 2c in https://pubs.rsc.org/en/content/articlehtml/2014/sm/c3sm52959b . If i recall correctly running time is roughly linear with the contraction/compression rate and for the compression rate 1e-2 for the LS algorithm (10 000 particles) it is ~2 minutes, for 1e-5 it is a couple of days (which roughly fits the linear scaling). Also, it is roughly linear with the number of particles. I.e. for compression rate 1e-2 for one million particles you will need ~200 minutes (but it is just a rough guess, also if my recollections are correct). I ran the algorithm up to 300 000 particles, it worked fine. A colleague of mine could run his own implementation of the Jodrey-Tory algorithm for one million particles (but this algorithm produces much lower densities for a comparable running time). I presume the program shall run fine for 1 million particles.

The algorithms here are not well parallelizable, so they run on a single CPU thread on a supercomputer as well, so generation will be roughly the same in speed or even slower then on your PC or laptop (depends on supercomputer CPUs). You can benefit from a supercomputer if you need to generate thousands or millions of packings, then you can run generation for all of them in parallel. Also, a supercomputer may have better or more suitable for the hardware compilers (IBM or Intel). But i don't remember if i tested the program with them. GCC is for sure a safe option.

Best,
Vasili