Computational time for mapdl.solve('ELFORM')

[lschuler1]

Hi,

I have observed that the command to reform the element matrices mapdl.solve('ELFORM') requires as much time as the mapdl.solve() command.
However, when issuing the ELFORM command, the problem must not be solved so I do not understand why the command is not faster.
Do you have any ideas on how to reduce the time needed to run the command mapdl.solve('ELFORM')?

Thank you,
Louis

[mikerife]

Hi @lschuler1 Louis. Can you please describe the FEM. How many elements/nodes are there? And what types of elements? Which MAPDL version are you using? How much time does it take to elform and solve?
When you say 'as much time' what does that mean? Generally speaking? Or did you time it somehow.
What physics and solver type are you comparing to?

There could be a pretty big difference in forming the stiffness matrix between a fully integrated 20 node hex (14 int points) or a reduced integration 8 node hex (1 int point). Something like that can throw off our expectations.

Or you may have found an issue that needs to be addressed. Are you an commercial Ansys customer? If so I'd recommend opening an Service Request. This seems to be an MAPDL issue that needs investigating and not specific to PyMAPDL.

If you cannot open an Service Request....well we will figure something out.

Mike

[mikerife]

Hi @lschuler1 & @germa89 Sorry, last night I had a brain cramp. Solve,elform only forms the full file for a perturbation analysis. If the analysis is not a perturbation, then a full solve is done even with the elform field set.

Please see the help on the WRFULL APDL command. It can be used to set the behavior to only form the full file at a specific load step. And the elform field of the solve command is not needed.

Mike

[mikerife]

@lschuler1 I forgot to add that you can look at the .dsp file for a finer break down of cpu/wall times for each step of the solve process (given per NR iteration). Plus other useful stats like OPSs to factorize vs OPSs to solve.

Oh and watch out for those small models. Of the 3 seconds reported to 'solve' probably 2+ seconds are accessing and writing the files to the hard drive.

Mike

[lschuler1]

Hi Mike @mikerife. Ok that explains why it was taking that much time. Unfortunately, WRFULL does not work for non linear problems. I will try to figure out something else.
Louis

[mikerife]

Hi @lschuler1 @germa89
Louis - I'm a little baffled by what the end game is here. It seems like you want to reach into the Newton-Rhapson algo and get out matrix/vector data as you need. At whatever load step, sub step, or NR iteration. This is quite the non-standard use...at least my gut says so. And it's been wrong plenty. Please open an Service Request and be prepared to describe in detail what you need to do. It may be possible for support to open a discussion with the appropriate MAPDL developer to get some answers for you.

Please feel free to send me any description directly at michael.rife@ansys.com I think that the old-fashioned method of writing a APDL solve control algo to solve models that have thermostats may be in-line with what you are doing and maybe a way to get out the data you need.

Mike

[lschuler1]

Hi Mike @mikerife ,
You are totally right, I would like to reach into the NR algo and get out some data. I know this is absolutely not the standard use, and that's why I have trouble to do it.
I'll send you an email with more details.
Thank you for your help.
Louis