How can I reduce computation time and memory to run this array file ? [FEKO V. 2023.1]

half_arc_a_46.98_final_probe_4by1_array_0.46_lambda_spacing_final_coordinate.cfx

Hi,

Currently, I am running this Feko file using MOM with parallel solver, but it is taking more than 48 hours to complete with continuous frequency points and fine meshing. Is there any hybrid solver you can suggest to me in FEKO to run this file with less memory and less amount of time ? [I need fine meshing and a continuous number of points].

Find more posts tagged with

Feko

v2023.1

Accepted answers

Torben Voigt

Hi @sswh2 ,

I used a standard mesh and then refined only the outer edges of the patches. I also set a minimum increment of 50 MHz for the adaptive frequency range to avoid too fine sampling and wasting runtime. Then I compared MoM and MLFMM. To make sure that the mesh was fine enough, I also refined it again. The results of the three models are practically identical:

Here are the computational requirements with 32 parallel processes:

Solver	Peak memory	runtime
MoM	3.868 GByte	1.678 h
MLFMM	5.752 GByte	0.851 h
MLFMM refined	7.686 GByte	1.068 h

You may download the files here: https://securefiletransfer.altair.de/link/OKIDvBRcfjvGtEmemq8JdM

Best regards,
Torben

Torben Voigt

Hi @sswh2 ,
I just reduced the wire radius (0.4 mm) to get rid of the warning about segment ratio. I'm sure you can change it to 0.65 mm again without significant changes in the results.

The difference between MLFMM.cfx and MLFMM_refined.cfx is the number of mesh elements. In MLFMM.cfx the vaule of variable mesh_refinement is "min(lambda/20 , Patch_L/9)", while in MLFMM.cfx it is "min(lambda/20 , Patch_L/7)". This variable is used to define the local mesh size of the outer edges of the patches:

Best regards,
Torben

All comments

Torben Voigt

Hi @sswh2 ,

Is this a duplicate of your post

"How to resolve the error 31027 for this file?[fEKO v 2013.1]" ?

Best regards,
Torben

sswh2

@Torben Voigt Sir, this post is not a duplicate. Previously, I had an error in the same file, so I posted it then. Now the problem is- simulation of this structure is taking more 48 hours and lots of memory - so I want to know if I use fine meshing and continuous frequency number of points then is there any hybrid solver available in FEKO which can reduce the computation time and memory? If you know how to run this faster, please suggest me.

Torben Voigt

Hi @sswh2 ,

Ok, understood. I will get back to you shortly. Note that you use MLFMM while the model is barely 3 lambda in size and in addition a fine mesh which is probably the cause for the high memory and runtime.

(MLFMM works best if a model is min. 4 lambda and the mesh is ideally not much finer than lambda/12.)

I will provide my recommendation at a later stage.

Best regards,
Torben

Torben Voigt

Hi @sswh2 ,

Here are the computational requirements with 32 parallel processes:

Solver	Peak memory	runtime
MoM	3.868 GByte	1.678 h
MLFMM	5.752 GByte	0.851 h
MLFMM refined	7.686 GByte	1.068 h

You may download the files here: https://securefiletransfer.altair.de/link/OKIDvBRcfjvGtEmemq8JdM

Best regards,
Torben

sswh2

@Torben Voigt sir, Thanks! First - which option did you use to refinethe edges of the patches only? second - I need my wire radius to be 0.65 mm for my design, not the 0.5e-4 mm, but you changed the wire radius in mesh. Last thing - in the solver settings between MLFMM and MLFMM refined, I can't see any differences. How can I select "MLFMM refined" in the solver settings?

Torben Voigt

Hi @sswh2 ,
I just reduced the wire radius (0.4 mm) to get rid of the warning about segment ratio. I'm sure you can change it to 0.65 mm again without significant changes in the results.

Best regards,
Torben

sswh2

@Torben Voigt Sir thanks for your explanation!