Re: [ESPResSo-users] Simulation of fluid using LB-method

[Wink, Markus]

Hello,

yes, you are right. I am getting 190k SUPS. 

The CPU it is running on is a Intel Xeon CPU @ 2.33Ghz, two processors in 
total. I hope that information helps since I am not familiar with CPU hardware.
The code is only running on one node since the mpirun-command is not working 
yet on my engine.

Best regards,

Markus Wink


-----Ursprüngliche Nachricht-----
Von: address@hidden [mailto:address@hidden Im Auftrag von Ulf Schiller
Gesendet: Montag, 27. Januar 2014 14:19
An: address@hidden
Betreff: Re: [ESPResSo-users] Simulation of fluid using LB-method

On 01/27/2014 02:07 PM, Wink, Markus wrote:
> Hello,
>
> I hope I understood your point correctly. I have run the script with two 
> calls of "integrate 1000", having a box of 50^3 and an agrid of 1, resulting 
> in 125*10^6 site updates.
> It took 22 min 2s for the simulation resulting in roughly 19k SUPS. This is 
> far below your expected 10M. I have no clue what the reason for that is.
>
> The binary of Espresso I have created only had "LB", "LB_BOUNDARIES" and 
> "EXTERNAL_FORCES" activated.  Running the testscript "lb_planar.tcl" (box of 
> length 12, total integrate of 5000) takes about 14 sec, resulting in ~0.6M 
> SUPS.

Hi,

well, Espresso has of course some overhead since it's not a pure LB 
code, but ~200k (that's what you get I think) resp. 600k seems pretty 
slow. What CPU are you running on?

Does anyone else have an estimate of how many site updates per second 
they get?

Cheers,
Ulf

> -----Ursprüngliche Nachricht-----
> Von: address@hidden [mailto:address@hidden Im Auftrag von Ulf Schiller
> Gesendet: Montag, 27. Januar 2014 12:16
> An: address@hidden
> Betreff: Re: [ESPResSo-users] Simulation of fluid using LB-method
>
> On 01/27/2014 11:54 AM, Wink, Markus wrote:
>> 3) Even though, the velocity profile saturates, it takes pretty long
>> (time scale of days). The saturation sets in at about 350 times the
>> "integration 1000" command. This corresponds to a time scale of ms in Si
>> units. Besides using OpenMpi or the GPU impementation, is there anyway
>> to speeden up the simulation? For example by adjusting the parameter tau
>> in the lbfluid command?
>
> Hi,
>
> that sounds a bit strange to me. If your system is 50^3 and you
> integrate 1000 steps, that gives 125*10^6 site updates. On a reasonably
> modern CPU, a reasonable LB code should get to around 10M SUPS (million
> site updates per second), that is, one call of "integrate 1000" should
> not take much more than 10 seconds, so 350 calls roughly one hour not
> days. Can you check how many site updates per second you get on your system?
>
> Cheers,
> Ulf
>


-- 
Dr. Ulf D. Schiller                        Building 04.16, Room 3006
Institute of Complex Systems (ICS-2)       Phone:   +49 2461 61-6144
Forschungszentrum Jülich, Germany          Fax:     +49 2461 61-3180