[feacluster]

I am trying to increase the mesh size of the pitzDaily tutorial so I can test performance of my cluster. In blockMeshDict I increased all hex blocks by a factor of 10. See below. The model runs, but crashes with mpi error after 7 iterations. It crashes at the same point on another cluster too. This lead me to think I am doing something wrong. Or is it some limitation of openfoam? I am running it on one node.

Any smoking gun before I share all details of system, compiler, etc..

[user@machine]$ diff blockMeshDict blockMeshDict.old
69c69
< (180 300 10)
---
> (18 30 1)
73c73
< (1800 270 10)
---
> (180 27 1)
77c77
< (1800 300 10)
---
> (180 30 1)
81c81
< (250 270 10)
---
> (25 27 1)
85c85
< (250 300 10)
---
> (25 30 1)

[Taataa]

Are you decreasing the time step with the same factor as well?

[feacluster]

Originally Posted by Taataa

Are you decreasing the time step with the same factor as well?

No, I have not changed anything else. Would that make a difference? Is it the endTime in the controlDict file I need to decrease by factor of 1000?

[einstein_zee]

Originally Posted by feacluster

No, I have not changed anything else. Would that make a difference? Is it the endTime in the controlDict file I need to decrease by factor of 1000?

Hi there,

the point is you need to maintain the same Courant number between the cases. And since you are refining the mesh you need to decrease your time steps also.

[Taataa]

Originally Posted by feacluster

No, I have not changed anything else. Would that make a difference? Is it the endTime in the controlDict file I need to decrease by factor of 1000?

You should try to maintain the Courant number. So if you are refining the mesh by a factor of two you should do the same for time step NOT the end time. So if for example if you time step is 0.001 and refine the mesh by a factor of two you should set the time step to 0.001/2=0.0005 assuming no changes in other parameters.

[feacluster]

Thanks for the replies. I changed the deltaT from 1 to 0.001 ( since mesh increased by factor of 10^3 ). However, it still fails after seven iterations:

Time = 0.006

smoothSolver: Solving for Ux, Initial residual = 0.776516, Final residual = 0.0514923, No Iterations 2
smoothSolver: Solving for Uy, Initial residual = 0.824286, Final residual = 0.0489998, No Iterations 2
GAMG: Solving for p, Initial residual = 1, Final residual = 0.0753774, No Iterations 6
time step continuity errors : sum local = 8.56395e+51, global = 6.169e+50, cumulative = 6.169e+50
smoothSolver: Solving for epsilon, Initial residual = 1, Final residual = 0.0774171, No Iterations 2
bounding epsilon, min: -9.23299e+99 max: 9.96454e+100 average: 7.49181e+94
smoothSolver: Solving for k, Initial residual = 1, Final residual = 0.000589475, No Iterations 1
bounding k, min: -8.51421e+86 max: 1.99993e+88 average: 9.75078e+81
ExecutionTime = 113.45 s ClockTime = 114 s

Time = 0.007


================================================== =================================
= BAD TERMINATION OF ONE OF YOUR APPLICATION PROCESSES
= PID 36271 RUNNING AT node12
= EXIT CODE: 8
= CLEANING UP REMAINING PROCESSES
= YOU CAN IGNORE THE BELOW CLEANUP MESSAGES
================================================== =================================
Intel(R) MPI Library troubleshooting guide:

[Taataa]

Refinement factor is 10 not 1000! So you should make it 0.1.

If it doesn't converge with reducing time step then assuming you've setup your case correctly, probably you need to make your tolerances tighter.

I'm afraid if you need more help then you need to provide more information about your case, BCs, schemes and solution.

[feacluster]

Originally Posted by Taataa

Refinement factor is 10 not 1000! So you should make it 0.1.

If it doesn't converge with reducing time step then assuming you've setup your case correctly, probably you need to make your tolerances tighter.

I'm afraid if you need more help then you need to provide more information about your case, BCs, schemes and solution.

Thanks, I changed deltaT to 0.1, but it still crashes in the seventh iteration. I am running the pitzDaily model. I assume this is some kind of universal benchmark? All I want to do is increase the computation time so I can test my cluster's performance.

Unfortunately, I have little background in CFD. More an HPC/FEA guy..

[feacluster]

Perhaps a better question is are there Openfoam benchmarks I can run to test mpi performance? For example, Ansys FLUENT has these benchmarks:

https://support.esss.com.br/hc/en-us...enchmark-Cases

[einstein_zee]

Originally Posted by feacluster

Perhaps a better question is are there Openfoam benchmarks I can run to test mpi performance? For example, Ansys FLUENT has these benchmarks:

https://support.esss.com.br/hc/en-us...enchmark-Cases

Hey there,

I just googled it and found useful information. You also might find your answer here.

How to test openfoam benchmark ?

https://openfoamwiki.net/index.php/B...ks_standard_v1