Re: [ESPResSo-users] lbfluid GPU vs CPU

[Dominic Roehm]

I tested it with reasonable viscosity and time steps and it was still present. I also changed the code to use double precision and it looked still the same. So I don't think that reducing the grid spacing would help and I have no clue what causes that behavior. But we had some issues with (artificial) oscillations in the GPU code in the past.

Dominic

On 06/25/2015 03:08 PM, Georg Rempfer wrote:

I'm sorry, I don't have time to answer those questions myself.

On Thu, Jun 25, 2015 at 3:08 PM, Georg Rempfer <address@hidden> wrote:

If you look at the actual vector valued oscillations, what's the pattern?

What happens if you increase the viscosity (say by a factor of two)?

What happens when you set agrid=0.5?

What happens when you change the time step?

On Tue, Jun 23, 2015 at 2:22 PM, Markus Gusenbauer <address@hidden> wrote:

Hi,

there is an issue with lbfluid comparing GPU and CPU. I have a rectangular channel with wall velocity inflow from the left. With CPU the parabolic flow looks fine, but with GPU there is a standing wave (see pictures attached). I have a minimal working example below.

What could be the problem? I tried version "ESPResSo-3.4-dev-172-g5e97dc1-dirty-git" and the latest master from today.

Markus




setmd time_step 1.0
setmd skin 0.2
thermostat off

set boxX 30
set boxY 10
set boxZ 10

setmd box_l $boxX $boxY $boxZ

lbfluid cpu grid 1. dens 0.008 visc 0.032 tau 1.0 friction 0.5

# fluid inflow
lbboundary wall normal 1 0 0 dist 1 type 501 velocity 2.5e-4 0. 0.

lbboundary wall normal 0 1 0 dist 1 type 501
lbboundary wall normal 0 -1 0 dist [expr -$boxY+1] type 501
lbboundary wall normal 0 0 1 dist 1 type 501
lbboundary wall normal 0 0 -1 dist [expr -$boxZ+1] type 501

############################################################################
# main iteration loop
set cycle 0
set endCycle 2000

while { $cycle<$endCycle } {
    puts "$cycle"
    if { $cycle == 1999 } { lbfluid print vtk velocity "fluid.vtk" }
    integrate 10
    incr cycle
}

-- 
   Dominic Roehm
   Dipl.-Phys.
   Institute for Computational Physics
   Universität Stuttgart
   Allmandring 3
   70569 Stuttgart, Germany 

   Office: 1.077
   Phone: +49(0)711/685-67705