[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: max number of CPU's

From: Jean-Noël Grad
Subject: Re: max number of CPU's
Date: Fri, 07 Jan 2022 18:37:47 +0100
User-agent: Roundcube Webmail/1.3.17


For domain decomposition, there must be at least two cells per direction, and at most 32 cells per direction (for a cubic box). The number of cells per direction depends on the interaction range. You can find the algorithm implementation in [1]; it solves up to 9 equations (3 for each direction), and there is not always a solution.

As a rule of thumb, for a cubic box without NpT, the number of cells per direction is often `floor(l_box / (l_cut + l_skin))`. For example, in a system with length l_box=12, LJ cutoff l_cut=2.5 and Verlet skin l_skin=0.4, the number of cells is 4 in each direction, which limits the number of MPI threads to 64.

For N-square, since the pair search algorithm considers all particles, there is no constraint from the interaction range. There are however special considerations regarding the number of MPI threads [2]. I'm only mentioning this for completeness, domain decomposition is usually the preferred cell system.



On 2022-01-07 02:34, Dusan Racko wrote:

please, is there a way to determine the maximum number of CPU's for a
parallel job
for a given box side length l_box, with LJ potential truncated at
l_cut and skin set to l_skin?
How can I estimate the number of processors based on these parameters?




Polymer Institute of the Slovak Academy of Sciences

Dubravska cesta 3
845 41 Bratislava, Slovak Republic [‡]
tel: +421 2 3229 4321



reply via email to

[Prev in Thread] Current Thread [Next in Thread]