What techniques can I used to optimize HFSS simulations?
Build A 9dB, 70cm Collinear Antenna
Both HFSS files have a problem in that the length of the section at the top of the antenna is too short, but that's a minor thing. One was produced by me and takes 4 hours to run on an 8-core 3 GHz Xeon Linux server. The other is a modification done by an Ansys engineer. That takes 14 minutes on my 2.2 GHz quad core Windows laptop. (He said it took 10-12 minutes on his laptop, so I guess he has a better laptop than me!)
But there's clearly a dramatic change in the run-times of these two files. I've been told the changes are too complex to discuss over email, and so I will have to visit them to go over the files, but I wonder if anyone here would like to take a look and advise me where the changes are.
The fast version uses a "fast" sweep, not an "interpolating" sweep as the slow version does. That was done to allow me to plot a graph of gain vs frequency. It might have some effect on the run-time, but there is a lot more than just the sweep type changed.
Does anyone have any material they can share on optimising HFSS simulations? I have so far been unable to get much from Ansys, although they clearly know how to do these things. I think I need to go on a more advanced HFSS course, but in the mean time, if anyone has any ideas, do let me know. (Obviously faster hardware would help, but 8 cores of 3 GHz Xeons are not slow).
Dave
Dave, which version of HFSS?
The files are HFSS14. i have tried looking into both the files.didn't seem to have any difference in the setup and mesh analysis.i have read some points on how to speed up the simulation process,long time ago.things like symmetry,defining materials properly and so on. cant seem to find the document,when needed.
I'm seeing HFSSDesign1 and HFSSDesign2 under 'fast' and 'slow' models ... there is a restriction under "Mesh Operations" -> Length1 on "Element Based Length Refinement" for restrict length of elements to length=80mm for each HFSSDesign2 ...
Dave, which Design ran successfully (and fast) - 1 or 2? (BTW HFSSDesign1 failed for me on 'fast' a little bit ago using ver 13 hfss)
Jim
JIM
from Dave's input.probably HFSSdesign2 in both are the one's to be looked at.excitations are another indication.they have been properly set in design2 with a PEC plate covering the waveport,so the waveport inside the radiation error doesn't show up.while for the design1 there are two excitations,which don't make sense.i have got an error in the beginning about the files being in version14
but in the first case'fast' length based mesh refinement is not enabled even though it is set,its not enabled but for the slow version it is enabled.i am running the simulation without any mesh refinement.i will post updates as soon as i have some.
ChachitoEL
Okay ChachitoEL, I noticed the difference in excitation on 'fast' #2 while the simulation is still running. fast #2 should be finishing for me shortly.
Jim
Thank You JIM. the maximum length in fast simulation is 247.8mm,which is over 3 times the max length of the meshrefinement upperlimit in the slow file.that could be the reason,just speculating here.got a gain of 7.62dB running the sweep now.simulation took just over 25 minutes to finish --> 'fast' file. my computer has i5 processor 64bit with 6 gig's of ram,hfss v13.
ChachitoEL,,
Solver is still running, but a 3D polar plot is available at this point (and it looks reasonable) after an hour on a 3 GHz Intel HyperThread (2) core, 2MB L2 Cache, 1 GB DDR2 RAM. Normally this is my LabVIEW platform, so not optimized for these killer 3D fullwave EM simulators!
Jim
Thanks again JIM for the post.
My 'slow' file is running right now.for me the big difference that is between these files is the 'MESHrefinement' without having any meshrefinement the fast file finished in 25+- time. with the meshrefinment in slow file i am still in the fourth adaptive pass,have a feeling that i might run out of Memory before the simulation converges,because the last pass took up almost all the available memory and so far its been 40 minutes and still solving adaptive passes.
i'm gonna call it off a night and check back here for updates tomorrow morning.
Have fun JIM
ChachitoEL
The files were generated with HFSS 14. Are people managing to open them in version 13? I assume at the very least that would give a warning, but is it opening?
Sorry, ignore HFSSDesign1 - that has an error in the port definition. I should have deleted that.
HFSSDesign2 is the one to be run, as some of you have worked out.
The slow file has run for 8 hours on an IBM server with a pair of quad core processors and 32 GB RAM. I think it's probably close to finishing, as the discrete frequency steps used range from 300 to 600 MHz. HFSS started doing them at 600 MHz, and is dropping the frequency each time. It is now on 330 MHz, so will probably finish in an hour. So I think my run times are:
- Slow file - 9 hours (540 mins) on 2 x quad-core Xeon processors running at 3.16 GHz (8 cores in total)
- Fast file - 14 minutes on a quad core laptop running at 2.2 GHz
I think the fast file is only stepping in 10 MHz steps, whereas the slow one is in 5 MHz steps, but that's only a factor of 2. The run times differ by a factor of around 39. And remember, one is a laptop, the other is a fast Xeon server.
I don't think the slow file would run on a laptop, as it is using over 10 GB RAM, and the laptop only has 8 GB. (It might work with swapping, but I might die of old age first).
---------- Post added at 02:03 ---------- Previous post was at 01:51 ----------
Yes, it is probably the mesh refinement. But the trick is to know how to refine the mesh. If anyone has any documentation about improving performance I'd appreciate it.
If one can change the run-time by an order of magnitude, it makes a huge difference in possible uses of the program. It would be a nightmare to try to optimise this antenna in any way by changing the lengths of the various bits if it takes so long to run.
That said, now I've run one simulation, I can see the peak gain is close to where I'd calculated it would be, so I'm not too far off. Before that, I had no idea, so was sweeping over a wide range. Now with a bit more insight, I can restrict the frequency range I scan over. But it would be practically impossible to gain much insight into the antenna when the run time is nearly 10 hours.
Dave
Hello Dave.
i went to bed with the simulation on.But the simulation never finished because of insufficient memory, these error was after two hours of simulation. so yeah i hope you have a better understanding of how things work.i wanted to break down the adaptive passes time and the sweep time and calculate the effictive time difference between the slow and fast file.But unfortunately doesnt seem like the slow file is can be simulated on this laptop of mine.
But what i have seen is sweep time on fast took like 4 minutes or so, where as on slow couldnt get there.
Adaptive passes took around 20 minutes and after the fourth pass it converged and this is the fast file.where as the slow file was no where close to the meshrefinment size of 80mm and was on the fourth pass and had a long way to go to converge after two hours.
Elchachito
I've now herd back from Ansys on the changes between the fast and slow files
The changes I made involve only the radius of your cable so feel free to
adjust number of elements and airbox size. You will also be able to run the
model at higher frequency but in reasonable time.
The thing is, the radius of the cable does not appear to have changed to me.
I can accept that changing the size of cable, if keeping it 50 Ohms, probably does not make much difference to the simulation since the cable is much shorter than a wavelength in radius and comparable to a wavelength in length, so one might expect using a larger cable to improve the simulation time. But the sizes appear unchanged to me.
But I've already been told this is too complex to sort out with email exchanges, so I'm not really in a position to keep asking Ansys for more information.
The slow simulation took 10.2 GB RAM on a Linux machine. That's the maximum amount of RAM reported with HFSS -> Solution data. I expect that's for the memory used by the solver, and probably excludes that used for the GUI. So I can well believe the slow version can't be run on a laptop or low-end machine. It took nearly 9 hours, though I can't quite work out from this below exactly how much time was taken. It appears to have started at 16:00:12, finished at 00:42:53, which is a difference of 8:42:41, yet the data below implies it took 08:38:09, which is a difference of over 4 minutes. I can't quite work out what these times refer to. I think the 38:20:14 is the CPU time used.
Start Time: 01/25/2012 16:00:12, Host skylark, HFSS Version 14.0.0
Executing from /opt/ansoft/hfss14.0/Linux/HFSSCOMENGINE.exe
Desired RAM limit not set.
Solution Basis Order: 1
Solution Sweep Discrete Sweep
From 0.3 GHz to 0.6 GHz, 60 Steps
<snip>
Solution Process Elapsed time : 08:42:40 , Hfss ComEngine Memory : 221 M
Total 08:38:09 38:20:14 Time: 01/26/2012 00:42:53, Status: Normal Completion
There's clearly a lot more to HFSS than just putting ones model in and waiting for it to finish.
Could you post the Mesh statistics from the solution data for the slow file.since i cant finish the slow simulation.it will be even better if the whole solution data can be uploaded here for a good comparison.
Thanks a lot Dave.
Regards
Elchachito
Soon after posting the run times HFSS locked up and I was unable to save the data, so I've lost the results from the 9 hour "slow" simulation. I'll run it again, but don't expect the results to be posted too soon.
Dave
Thanks dave.
Post them whenever you have them up. thanks a lot one more time.
Regards
Elchachito
optimize techniques simulations 相关文章：