No, I set a skin of 0.1. I think I found the problem. The max_range is 3.4, while when I use 16 cores, node_grid is 4 2 2, so on one box size it's 20/4=5 < 2*3.4, which cause the problem. Thanks. Xikai
Date: Thu, 17 Jul 2014 23:14:41 +0200 From: address@hidden To: address@hidden; address@hidden Subject: Re: [ESPResSo-users] minimum number of cells
On 17.07.14 20:23, Xikai Jiang wrote:
Dr. Arnold:
Yes, it's an error, not warning, for "number of cells is smaller
than minimum". The simulation box sizes are 20, 20 and 19 and
each side is larger than twice the r_cut. P3M parameters are
from automatic tuning.
That sounds weird, since r_cut is much smaller than the simulation
box. Do you set a large skin value? Otherwise, there should be
around 6 cells per dimensions, so you can use up to approx 200 cores
before getting the error.
It requires at least 2 cells per spatial dimension,
then "min_num_cells" is 8 on 1 core, and it's 1 on 8 cores. For
16 cores, if I set min_num_cells to 1, for the whole system, the
total min_num_cells is 1*16=16 and it's safe for the simulation,
is my understanding correct?
On 16 cores, min_num_cells is anyways 1, as you can check using
setmd. Check also what max_range, that should be p3m skin + r_cut,
or whatever else has the largest cutoff. If this cutoff is larger
than box_l/num_procs_per_dimension, then you get a problem (or
box_l/2 on a single core).
Axel
Date: Wed, 16 Jul 2014 22:12:01 +0200
From: address@hidden
To: address@hidden; address@hidden
Subject: Re: [ESPResSo-users] minimum number of cells
On 16.07.14 18:59, Xikai Jiang
wrote:
Dear all:
I have a question about the global variable
"minimum_num_cells".
Espresso seems to set "minimum_num_cells" to 8 by default,
but
when I use MPI and use more cores, it gives me a warning
saying
"number of cells is smaller than minimum". And I found
when I
increase number of cores, number of cells
decreases.
You don't get a warning, but an error message, since Espresso
would produce incorrect results. We need at least two cells
per spatial dimension to make the minimum image convention
unambiguous. minimum_num_cells however counts per core, so if
you have 2 cores, you just need 4, and for 4 cores 2.
I'm using P3M algorithm and Coulomb cutoff is 3.3:
inter coulomb [expr 138.935485/$epsilonr/$T] p3m 3.3
100 7 1.1
Here I'm supposed not to change the interaction range
(r_cut) and
my question is, is it safe to manually set
"minimum_num_cells" to 0,
in order to avoid the warning and let me use more cores?
No, first of all you cannot set minimum_num_cells to a lower
value, since ESPResSo wants to avoid computing incorrect
results. As a consequence, there is a limit to the cutoff
depending on the box size. You simply cannot use r_cut 3.3 in
a box smaller than 6.7. Either, you put more particles to
increase the box size, or you don't insist on your parameters,
but rather let P3M tune them.
Best,
Axel
--
JP Dr. Axel Arnold
ICP, Universität Stuttgart
Allmandring 3
70569 Stuttgart, Germany
Email: address@hidden
Tel: +49 711 685 67609
--
JP Dr. Axel Arnold
ICP, Universität Stuttgart
Allmandring 3
70569 Stuttgart, Germany
Email: address@hidden
Tel: +49 711 685 67609
|