Memory issue

I want to simulate a hole array filter but I'm having trouble with memory usage. At first I would get an error saying insufficient memory available so I made the mesh much more coarse but it still took over 24 hours to run the solver. Any suggestions?

Unable to find an attachment - read this blog

Find more posts tagged with

English

Feko

Accepted answers

All comments

Torben Voigt

Could you attach the *.out file here?

It seems that something is slowing down the simulation at the 6th frequency. We're currently investigating.

Altair Forum User

Could you attach the *.out file here?

It seems that something is slowing down the simulation at the 6th frequency. We're currently investigating.

Thank you so much for taking the time to do this.

Unable to find an attachment - read this blog

Mel

Hi Pavlam

There seems to be a slow down problem with this specific model.

On my machine it runs 46x faster with FEKO version 2017.2. (Note I only tested frequency=900 GHz)

Torben Voigt

Hi Pavlam,

The reason for the runtime issue is still under investigation. It could be worked around by setting up the unit cell differently:

image.png.6172c731c3723a5133030ca21e26ed0a.png

Note that your mesh setting of lambda/5 does reduce the memory requirement, but will most likely not lead to trustworthy results. The unit cell has a size of 1.6 x lambda and thus will demand some memory with a mesh size of lambda/10 (>20 GByte). I see from your *.out file that you have 64 GByte installed, so that should not be a problem.

BUT: Due to the size of the unit cell there are grating lobes for each frequency (Grating Lobes). They prevent an accurate calculation of the transmission / reflection coefficients. A simple workaround here is shown here:

image.png.2ad49e14550b3aeaf371f392a220cc4d.png

Note, P_incident is calculated by 2nd model, with no geometry present, just calculating the power going through NF front.

I will add the files here once the simulation has finished.

JIF

Regarding grating lobes, maybe also read

Torben Voigt

The files are now attached. Looking at the results I have to admit that they could look better (Rx+Tx =! 1 here and there). Most likely this can be improved by a larger distance between unit cell and near field requests. I chose 1 lambda at highest frequency, but looking at JIF's link, 10 lambda at lowest frequency probably would have been much wiser.

<?xml version="1.0" encoding="UTF-8"?>

Unable to find an attachment - read this blog

Altair Forum User

Hi Pavlam

There seems to be a slow down problem with this specific model.

On my machine it runs 46x faster with FEKO version 2017.2. (Note I only tested frequency=900 GHz)

Yes, 2017 definitely ran it faster. Thank you for the recommendation it helped a lot

Hi Pavlam,

The reason for the runtime issue is still under investigation. It could be worked around by setting up the unit cell differently:

<?xml version="1.0" encoding="UTF-8"?>

This geometry was much simpler and easier to run. I would have never thought of this. Thank you for taking the time to do all this, I really appreciate it.

Torben Voigt

Hi Pavlam,

One imoprtant addition: The calculation the Rx / Tx coefficients from the power flow through near field is only well-defined if the size of the near field requests is N * size of unitcell (N is an arbitrary interger value). In your case the size could be dy*dz or dy*2dz,...