HFSS eats up memory - help needed

[paszczakojad]

HFSS eats up memory

Hi,

I'm optimizing a waveguide filter using HFSS 11 and I noticed two odd behaviours:

1. More serious: each iteration eats up some RAM. After 100-130 iteration all 3.2 GB of RAM (I'm using XP 32-bit) is used up and HFSS crashes. After restarting everything works fine (I'm using Quasi-Newton optimizer, so restarting it after an appropriate number of iterations still provides some progress in the optimization process), but that's annoying. When I have an urgent project I have to restart HFSS every 6-8 hours (including waking up in the middle of the night :[ ). I tried different memory limits in Tools->Options->HFSS Options, but that didn't help. Can I fix that somehow? I don't think the optimizer (at least Q-N) needs to have hundred of results in RAM. And even if it needs so - the parameters and cost should be enough, that doesn't need too much memory. I suppose it stores all the meshes in the memory (the 'Save Fields' box in Analysis->Setup1->Sweep1 is unchecked). I really dream about leaving the optimization for a week unattended...

2. Less serious problem: each interation is slower than the previous one - after say 100 iterations HFSS waits for a really long time after each iteration displaying messages like 'updating all derived data centrally'. The cure is to stop simulation, clean up solutions (HFSS->Results) and restart simulation - then it goes smoothly for some time. Restarting without cleanup doesn't help. I tried to check 'Apply variation deletions immediately' in 'Tools->Options->HFSS Options', but that didn't help. I could fix the first problem by using x64 OS and lots of RAM, but after hundreds of iterations the optimization process will be painstakingly slow anyway. Any suggestions?

P.

 

[hughrpg]

Of course each iteration uses more ram.. it is using ever more tetrahedra to mesh.

Waveguide filtters are tricky, as they are large compared to lambda but there are some get-arounds:

If it is a normal type of filter, these are very well understood, and there are standard formulae for calculating a starting point.

They are usually symmetrical in more than one axis: you can use Symmetry planes to drastically reduce the size of your model (and speed up the solve time!)
You need to of course set the impedance multiplier correctly. There are relatively good documents how to do this in HFSS, and previously (version 8 days) was very standard, as CPU power was a fraction of todays.

Use the "assign mesh operation" to manually mesh critical areas before you start.

Using the optimiser in HFSS is rather optimistic. Much better is to use some other way to design the waveguide filter, and then test the results in HFSS and make final adjustments if necessary. A week of optimisation is quite a long time

If you really want to make your own fancy type of waveguide filter
There are far better ways to do this.. for example using mode matching software (see Mician uWave Wizard) which can optimise you a near perfect solution in minutes.
You can download an eveluation version, which I am sure will get you extremely close to the final result far faster than HFSS.

>after say 100 iterations HFSS waits for a really long time after each iteration >displaying messages like 'updating all derived data centrally'.

I am not surprised... it is shovelling around a LOT of data.

Hope my comments are helpful.

 

[katko]

Ever used HFSS for a decent-sized project? My 2GB RAM PC has crashed after 20 iterations on smaller models, let alone 100. Use symmetry, reduce size if possible. Do you really need over 100 mesh refinements? What's your convergence criterion?

 

[paszczakojad]

To hughrpg: thanks for your hint about uWave Wizard. I have to design an iris diplexer, so that seems to be a perfect choice

To katko: I meant 100 optimization iterations, not adaptive mesh refinements.

Anyway - I'll give a try to uWave Wizard now

Thanks,

P.

 

[rohanapp]

Regarding progressive slowdown of optimization: It could be due to the reports. You may want to delete all your results (reports and field overlays) and see if the issue goes away.