yade-users team mailing list archive

["Christoph Kloss" <christoph.kloss@xxxxxx>]

Hi Feng, dear yade community,

>>Is it correct that Yade is based on a shared memory approach?
>>Vaclav has answered that problem, the performance increase on the parallel
issue using OpenMP seems to be limited.

Yes, especially since shared memory machines tend to be costly with the
large number of CPUs required for
high end computing. Modern clusters have shared/distributed memory
architecture so that you can use one node
with, say 8 or 16 procs in shared memory mode using OpenMP and as many nodes
as you like using MPI.

>>2. The maximum particle number possible on a cluster
>>Feng:In my thesis I used 8200 particles on a workstation, I wish I could
use at 10^6 level.
With my EDEM-FLUENT coupling, I used up to 700,000 particles but performance
was quite poor at this
number of particles.
With "LammpsFoam", it should be possible to dramatically speed this up by
factor 5-10 or so, so that 1e6 particles should be 
no problem.


>>5. Is it easy to change/create new particle constitution laws for the
solid phase?
>>Feng: I ask this because I am working on visco-elastic model (Burgers,
Maxwell etc which can be used in asphalt concrete) 
>>      and some other non-linear models

Well, if you are referring to solving a continuum-based granular model:
LAMMPS has a (I think quite simple) Finite Element Code included, see
http://lammps.sandia.gov/doc/fix_atc.html 
Actually, this is also a topic we are working on, coupling continuum and
discrete granular models.
I think a promising strategy could also be to use e.g. LAMMPS for the
particles and OpenFoam for the continuum model

I am happy that some discussion started off. I think we should keep in
contact to exchange thoughts and ideas.
I will be visiting Particles2009 in Barcelona, CFD2009 in Melbourne and
WCPT10 in Nuernberg. It would be nice to
meet some of you there.



Regards,

Christoph


=========================

Von: Feng Chen [mailto:fchen3@xxxxxxx] 
Gesendet: Freitag, 09. Oktober 2009 05:07
An: kloss@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx;
yade-dev@xxxxxxxxxxxxxxxxxxx; yade-users@xxxxxxxxxxxxxxxxxxx
Betreff: RE: [Yade-dev] Open Source DEM with LAMMPS

Hi, Chris:

Thanks for your answer, I would like to share my experience on the issues
below:

Btw: Is it correct that Yade is based on a shared memory approach?

///Vaclav has answered that problem, the performance increase on the
parallel issue using OpenMP seems to be limited.

>1. FVM size/particle size

The simulations I did so far were with particle diameter in the same order
of
magnitude or smaller than the grid spacing.
I did not extensively test particles that are much larger than the grid size
because my flow problems did not involve such cases.
But to my knowledge, this would lead to a) convergence issues and b)
possibly malpredictions of the dragforce by standard drag correlations
(Gidaspow, DiFelice,...)
My maximum ratio cell spacing/particle diameter was ~1/1 which converged and
matched the validation experiments well

///As I know the minimum FVM size must be larger than the particle size,
when the FVM size is close to particle size, will lead to porosity value
singularity, which will in turn, cause instability in N-S continuity
solution (This is also mentioned in PFC2D fluid option's manual)


>2. The maximum particle number possible on a cluster

It depends on what cluster you have :-)
If you look at http://lammps.sandia.gov/bench.html#chute
you fill find the rightmost dot in the left figure corresponding
to 2 BILLION PARTICLES!

///In my thesis I used 8200 particles on a workstation, I wish I could use
at 10^6 level.


>3. The parallel mechanism

We are using MPI because we think that the code should be portable to
clusters.
We have a small cluster (~150 proc.) at our Institute, so for me it is a
huge speed-up
switching from my workstation to the cluster

///Same answer as the "shared memory", I met the "bottle neck".


>4. What kind of problem/type of flow you use in your problem?

Typical problems that we are using or intending to use CFD-DEM for is
pneumatic conveying,
fluidized bed, hopper discharge and die filling.

///I use this in the chromatography column packing, stress-strain caused by
fluid drag, sheet pile, seepage in dam


>5. Is it easy to change/create new particle constitution laws for the solid
phase?

Yes, LAMMPS is written in the intention to be easily modified. A whole
section of the documentation is dedicated to this:
http://lammps.sandia.gov/doc/Section_modify.html
I must admit that currently, LAMMPS is missing a feature to import granular
walls from
CAD. But I am working on this...

///I ask this because I am working on visco-elastic model (Burgers, Maxwell
etc which can be used in asphalt concrete) and some other non-linear models


>6. Effect of different drag equations

I must admit that I did not extensively test that. Afaik, Gidaspow could
lead to convergence problems as it is discontinuous in one point. Thus I
also used
DiFelice but did not find much difference

///I used Ergun's (1952) equation, there is a more detailed comparison in
MFix about different drag equations

I should add that the simulation experience I am talking about was with a
EDEm-FLUENT coupling that I wrote myself. The "LammpsFoam" is now in beta,
and this fluidized bed configuration
http://lammps.sandia.gov/movies/gran_openfoam.avi
is the only case that we calculated with "LammpsFoam" so far.

///YADE-OpenFOAM has also been in an "alpha-beta" status for a long time,
some portion of the code (especially coupling part) is so crappy that I even
feel shame to release it:-P, although I have been working on this for 2 more
years, hehe :-)