long simulation time issue

Marcus Chang

Dear support team,

 

Recently, we run a simulation for a absorber placed between sphere two horn, we would like to get S parameter from this setup.

However, this simulation was stuck as shown in attached figure, LU decomposition of the matrix, for almost two days.

And also I think our server system is good enough for this simulation with 512 RAM.

In this simulation, how can I decrease the simulation in hour through the setting, I think it is must be something wrong in my setting for absorber in this simulation.

Because it is one of an urgent studies for a paper, pleas help us solve this issue ASAP, appreciated.

 

BR/Marcus

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

 

Marcus Chang

Dear all,

 

So far I try to modify solver setting, active “Solve model with the multilevel fast multi pole method (MFLMM)”, and it’s seemed much more faster now.

If there is any further question, I will let you know, thanks a lot.

 

BR/Marcus

Marcus Chang

Dear all,

 

So far I have two questions shown as below, please check:

 

1. Is it possible to reduce the simulation time by modifying the setting, the .cfx file is shown as attached.

Next step we will add a reflector with a size slightly larger than the sphere, and for this reflector, we need to take diffraction issue into account.

For this kind of problem, if we add the reflector described as above, do you have any better suggestions for us to keep well accuracy and also spend less time to simulate it?

In this scenario, we need to rotate both TX/RX horn per two degrees in a circle (simulate 181 times by using parameter sweep in FEKO) and reflector included,  we spend 5 hours for each simulation, therefore, we will spend 181*5 hours, 37 days to solve it by using workstation with 56 processors CPU (Intel(R) Xeon(R) W-3175X CPU @ 3.10GHz) and 512G RAM.

The total simulation time and more details for one of simulations are shown as below from the end of .out file:

 

In this case, the solver for reflector we use is PO, however, as mentioned above, we need to take diffraction issue into account, but if we use MoM/MLFMM method, the simulation time will be much more longer than using PO method. For this scenario, please provide us any suggestions to avoid this issue.

 

2. There is a weird results in S-parameter for attached .cfx file, that I have no idea is, there is a little difference between S21 (-85dB) and S12 (-90dB) in such symmetrical geometry model.

And also, these two S-parameter values are lower than it simulated with two horns only (S21&S12=-79 dB), it doesn't make sense, the S21 values for TX/RX only should be the lowest when compared with the other scenario if we add additional objects. Even though the value for this difference is relative small, kindly give us explanation for this deviation, such as, it is caused by asymmetrical meshing or any mistake in my simulation setup, etc.

 

P.S: Due to the reflector we design is confidential, attached .cxf file we provide is not including reflector.

 

 

                    SUMMARY OF REQUIRED TIMES IN SECONDS

                                             CPU-time       runtime
 Reading and constructing the geometry          2.326         2.326
 Checking the geometry                          1.581         1.580
 Initialisation of the Green's function         0.000         0.000
 Calcul. of coupling for PO/Fock            12030.895     12030.895
 Transformation to equivalent sources           0.000         0.000
 Ray launching/tracing phase of RL-GO           0.000         0.000
 Calcul. of the MLFMM transfer function         0.483         0.484
 Fourier transform of MLFMM basis funct.        2.414         2.413
 Calcul. of matrix elements                   139.400       139.399
 Calcul. of right-hand side vector           8747.818      8747.818
 Preconditioning system of linear eqns.        18.643        18.643
 Solution of the system of linear eqns.       423.926       423.927
 Eigensolution for characteristic modes         0.000         0.000
 Determination of surface currents              0.000         0.000
 Calcul. of impedances/powers/losses            0.020         0.021
 Calcul. of averaged SAR values                 0.000         0.000
 Calcul. of power receiving antenna             0.000         0.000
 Calcul. of cable coupling                      0.000         0.000
 Calcul. of error estimates                     0.000         0.000
 Calcul. of electric near field                 0.000         0.000
 Calcul. of magnetic near field                 0.000         0.000
 Calcul. of far field                           0.000         0.000
                            other               1.593         1.594
                                         ------------  ------------
                         total times:       21369.099     21369.100
               (total times in hours:           5.936         5.936)

 Specified CPU-times are referring to the master process only
 Sum of the CPU-times of all processes:    598334.848 seconds (   166.204 hours)
 On average per process:                    21369.102 seconds (     5.936 hours)

 Peak memory usage during the whole solution: 930.652 MByte
 (refers to the master process only)
 Sum of the peak memory of all processes:     21.454 GByte
 On average per process:                      784.597 MByte

 

BR/Marcus

Marcus Chang

Dear FEKO support team,

 

Due to the paper submission deadline is coming soon,

my I have your response within this week?

I’m sorry if this has caused you any inconvenience.

 

BR/Marcus

Mel

Regarding 'there is a little difference between S21 (-85dB) and S12 (-90dB)'

This is because of very minor differences in the mesh. If you set geometric symmetry between the two horns the S-parameters will be identical.

 

Regarding the solution, any reason why you are using PO?

A pure MLFMM solution takes roughly 20 GByte, a couple of minutes, and all effects will be taken into account.

 

Marcus Chang

Dear Mel,

 

Thanks for reply.

 

To clarify the problems we face, please see attached .cfx (similar to our case), in such simulation, we spent several hours to simulate it.

Is there any suggestions?

Could you please help us modify the setting if you have any better idea?

Thanks a lot!

 

BR/Marcus

 

 

Mel

Please attach the *.out file too.

Marcus Chang

Dear Mel,

 

Please see attached .out file run in our sever with 256 RAM.

Even though we spent 1 hrs to solve it faster than our expectation.

Actually, another simulation case (confidential) we run, we spent 7 hrs to solve it.

The difference between these two simulations are the location of the horn antenna pair, sphere and the strcture of the reflector (smaller structure compared to the file I give you).

I also attached this .out file, please check.

Thanks a lot.

 

BR/Marcus

 

Mel

Hi Marcus

 

Previously you stated:'To clarify the problems we face, please see attached .cfx (similar to our case), in such simulation, we spent several hours to simulate it.'

Today you stated:'Please see attached .out file run in our sever with 256 RAM. Even though we spent 1 hrs to solve it faster than our expectation.'

The *.out file shows a 1.1 hour solve time. Is this too long?

Note that for S-parameters if you do not need S12 and S22 you can set the 2nd port on the S-parameter request to inactive. This will save you 1470.531 seconds (see out file) for the 2nd solution. The 2nd port will only act as a sink port then.

 

Regarding the hallow sphere, I see stabilized MLFMM is activated. Note that it only applies to the metallic parts of the model and will run longer than for the case where stabilization is not active. 

 

Marcus Chang

Hi Mel, 

 

Even 1 hour, it is still too long, because in order to get a polar plot-like with 360 degrees S21 patterns, we need to rotate horn antenna pair per 2 degrees and run 181 times.

So we need to spent 181*1.1 hours to get the results, even more, for the another confidential file, (181*3 hours=22 days).

In real measurement, we will let 2nd port on.

Is there any other solution to save simulation time? Such as modify the mesh or some simulation setting we neglect, etc.

 

BR/Marcus

 

Mel

Hi Marcus

 

Is the reflector and sphere in a static position? So only the horn and absorber is moving?

Marcus Chang

Dear Mel,

 

Please see below screenshot link, only antenna pair rotate.

https://drive.google.com/file/d/1CAzEFV7QyJi0L9x0kIxtEUOVJSroXHiy/view?usp=sharing

 

Sweep = 2 degrees

 

BR/Marcus

 

Mel

You will need to do a full wave solution, so either MoM/FEM/MLFMM.

 

Some suggestions:

1 Do a mesh convergence test on the reflector and target to determine what the coarsest mesh is you can use.

You can vary the mesh sizes across the different parts/faces by using local mesh settings. I would suggest trying a mesh size of lambda/6.5 or maybe coarser. 

 

2 It should be possible to replace the absorber with an equivalent cuboid of the same width and length but with some thickness inbetween the cone base and cone tip. The cones individual surface areas add many expensive dielectric triangles to the model. A single cuboid with width/length/height will be computationally cheaper.

 

3. Instead of running every model with all the available cores, rather run 2 models over half the angles with half the cores each. The model is not very large and the parallel efficiency is not perfect for increasing number of processes.

Marcus Chang

Dear Mel,

 

Related to your suggestion 2, do you mean we can simplify the structure of the absorber to be a cubic as shown in attached .cfx file?

We finally get the results of the S21 and compare these two simulations w/ and w/o modification, we found that if we did the simplification, the S21 (-65dB) is 20 dB higher than original one (-85dB). 

 

 

BR/Marcus

Mel

Marcus

Your results make sense. Clearly the suggestion to replace the absorber with a cuboid to save on mesh elements is not suitable in this case.