Very long calculation time for casepump testcase on Tesla K80

Basti

Hello everyone,

I recently got access to a Tesla K80 GPU and wanted to give the "casepump" another try, since it took ages on the GPU of my working PC. However the output shows, that the simulation will run for 3 weeks to calculate 6 seconds of simulation time.

I feel, that this simulation with 1,000,000 particles shouldn't take that long. Could it somehow be the case, that the time step is way to small in my simulations?

DtIni=1.97385e-06
DtMin=9.86924e-08

is shown in the output. And it takes about 151990 steps for 0.06 seconds of simulation.

I didn't change the Def.xml.

Thank you alot in advance!

Alex

Yes! CaseDump is very very long.

Time step si too small, that is why many number of steps are required to complete the simulation. Why?
Because the time step (dt) is adjusted using the value of velocity of sound (Cs).

The speed of sound is computed as Cs=coefsound * speedsystem.
Where speedsystem is the maximum velocities expected in the case.
Here, the maximum velocity is the maximum linear velocity during rotation of the pump. In this case 175.5 m/s.

So we need to compute the correct speed of sound in the XML with:

- speedsystem value="175.5" auto="false" comment="Maximum system speed (by default the dam-break propagation is used)" BUT HERE WE INDICATE THE VALUE

- coefsound value="20" comment="Coefficient to multiply speedsystem"

- speedsound value="0" auto="true" comment="Speed of sound to use in the simulation (by default speedofsound=coefsound*speedsystem)" SO THAT SPEED OF SOUND WILL BE 20*175.5

Time step will be of the order of h/Cs... so very very small
We need a small dt to solve properly the scales of this simulation

Regards

Basti

Hello Alex,

Thank you very much for that explanation. Because of the high rotation speed of the pump this does totally make sense to me.

Thanks and best regards.