|
From: | Axel Arnold |
Subject: | Re: [ESPResSo-users] Increase the lattice boltzmann agrid but got error. |
Date: | Thu, 28 May 2015 23:11:45 +0200 |
Hi!
In fact, it is probably more safe to let the automatic tuning do its work, including the mesh. The rule of thumb is that there should be around 1 particle per mesh point, so 42 is way too large. This indeed can lead to strange behavior like sharp contrasts in the charge distribution, which leads to numerical errors. So, better rely on "p3m tune accuracy 0.01" or so.
The skin determines when the Verlet lists are rebuild and particles are exchanged between processors. A large skin means less communication and calculation of Verlet lists, but more comparisons in the inner particle loop. The optimal value there is a trade off that depends on your hardware (network latency vs. computing power). Fortunately, the computation time as function of the skin typically has a broad minimum, so that you don’t need to excessively tune. As a rule of thumb, the skin should be about 40% of the primary particle diameter. As I said, there are some side constraints: skin must be smaller than agrid/2, and processor domain size minus skin limits the range of short ranged interactions. Because P3M with very small grids tends to use rather large cutoffs, you can hit that limit. However, that was definitely not your case with 42... Best, Axel
-------------------------------------------- Axel Arnold
Martha-Schmidtmann-Str. 7 70374 Stuttgart, Germany Email: address@hidden Phone: +49 173 870 6659 |
[Prev in Thread] | Current Thread | [Next in Thread] |