Learning Forum

[lsmontal]

Hi,

I am running LES until convergence without data sampling, then engage data sampling for time statistics (sampling interval set to 100). However, the simulation becomes extremely slow when I do so. I only need time-averaged data over a portion of my domain, so I'm thinking about 2 approaches but am not sure how to do either:

1. Can I create a mesh zone based on coordinates and select this when setting up data sampling for time statistics?
2. Can I define a function to continuously export e.g. a text file with time-averaged data at each node within a region (instead of including this data in the .dat exports)

As a note, I was previously exporting instantaneous ASCII data over several planes and averaging in postprocessing, but this was also extremely slow + data-intensive.

[Rob]

It shouldn't have too much of an effect. Can you check RAM usage and disc space. 

[lsmontal]

 

Ok, thank you. The disc space is fine. Increasing the allotted memory did seem to help somewhat (I’m running this on a computing cluster), but it’s still much slower than before. Any other ideas?

FYI, I went from 200MB to 1GB per core, which looks excessive (~15% of memory used)… so increasing futher won’t solve it, it seems.

 

[Rob]

How many cells are on each core (roughly)? If you're only using 150MB per core you may find the message passing between cores to slow the solver down more than the extra cpu speeds it up. Unless I'm trying to use a single box here (28 or 32 cores) I tend to aim for 100k - 2M cells per core, it's a good trade off between cpu gain and not being greedy with core count. 

[lsmontal]

Hm, interesting. There are ~400k cells, and using 8 cores instead of 16 does speed it up. (4 is a bit slower, though.) Thank you, that's helpful!

[Rob]

You're welcome. Some of the newer chips are very quick, but don't have the memory channels capable of handling all of the traffic. Great for emailing while watching Youtube etc, not so good for many core simulations.