Learning Forum

[NimaNZ]

Hi everyone

When I run LES model on 2 nodes (2*20 cores each) re instantaneous results of flow characteristics are noisy with different trend compared with the time I run them on 1 node.

I’m wondering if is is something usual or there is something else wrong with my model?

appreciate any help

[Rob]

There shouldn't be a difference in the result with differing core counts, but can you check convergence as that differ if the mesh isn't suitable.

[NimaNZ]

no it goes to 1e-07 for continuity and 1e-9 for momentum. The mesh is okay I think.

[Rob]

Please post images of the results.

[NimaNZ]

,I am not near my laptop now. I have the graph of the lift coefficient through 1 second of the flow with 1 node and 2 nodes. (8s to 9s with 2 nodes). As you can see the second one is noisy.
another difference between them is that the first case was prepared using in-house HPC with 20*2 cores. But the second case was prepared using my laptop (6 cores). Can it be a problem? because I should set the number of cores on my laptop to 80 to be able to submit the job on HPC. Then other 80-12 cores data will be run on memory, right?
These noises are not effective in mean values; the will be problems in RMSE graphs.

[Rob]

Repartitioning when changing core count is automatic, don't worry about that. The noise on the second plot implies either poor convergence or that you're picking up some fast transients. Hence needing to see results etc. If you stretch the second plot out how does it compare: always size and scale plots to make it easier to compare data.

[NimaNZ]

thank you for your response. The convergence and mesh in both cases are the same. In terms of results, I have mean values validated, but there is no RMSE profile provided. Sorry about the plots, you're right about the scaling and comparison but it is still noisier. Just wanted to make sure if there is any technical reason about running on two nodes, if that's not the case then it will be fine.

[Rob]

Shouldn't be. We have seen some differences where the convergence & mesh were very poor but otherwise expect to see the same results.

[NimaNZ]

Thank you so much