Harmonic analysis using the frequency-sweep Krylov method - problem when running with nproc = 2+

[dhabar95]

When running the code (https://mapdl.docs.pyansys.com/version/stable/examples/extended_examples/Krylov/krylov_example.html):

mapdl.run('/SOLU')
mapdl.antype('HARMIC')  # Set options for harmonic analysis
mapdl.hropt('KRYLOV')
mapdl.eqslv('SPARSE')
mapdl.harfrq(0,1000)    # Set beginning and ending frequency
mapdl.nsubst(100)       # Set the number of frequency increments
mapdl.wrfull(1)         # Generate FULL file and stop
mapdl.solve()
mapdl.finish()


dd = mapdl.krylov       # Initialize Krylov class object

Qz = dd.gensubspace(10, 500, check_orthogonality=True)
print(Qz.shape)

Yz = dd.solve(0, 1000, 100, ramped_load=True)
print(Yz.shape)

result = dd.expand(residual_computation=True, residual_algorithm="l2", return_solution = True)

dd.gensubspace is unable to locate the file.full ( as i am using nproc = 4 and create fileN.full files).

Solution tried and failed:

1. Using COMBINE - doesn't work for hropt("FULL")
2. Using DMPOPTION before mapdl.solve() - doesn't combine the fileN.full files
3. Using nproc = 1 - Takes ages to solve

Could you please help how to solve this problem about nproc and file.full problem?

[mikerife]

Hi @dhabar95 @germa89 @clatapie
In MAPDL there are two Krylov workflows; one that uses APDL commands and one that uses a set of 3 APDL macros (that have to be downloaded from the MAPDL documentation). The APDL macro method notes that SMP must be used. I took a quick look at the APDL macros vs the PyMAPDL Krylov method and kind of gave up for now (a lot to look at). So instead in the example Dharshan is pointing to I used:

mapdl = launch_mapdl(nproc=4, additional_switches = '-smp')

And the example ran fine. I don't have a lot of time to dig into the APDL macros and PyMAPDL Krylov method today...maybe later this week. I'm not yet sure why SMP is required.
Mike

[germa89]

Thank you for this @mikerife

I think there was some limitations regarding this on MAPDL side.... But I'm not sure. This could be solved easily though.
However we do test now in DMP mode in the CICD #2308, we should have caught this error.

Pinging @kmkoshy and @FredAns for further input.

[germa89]

Linking the other answer here too: #2306 (comment)

[FredAns]

Hi @germa89 @mikerife @kmkoshy ,
KRYLOV macros are based on APDLMath commands, and we are in the process on having these commands fully compatible with DMP MAPDL, but we are not done yet. That's why we still need to run this feature in SMP mode.
Thanks

[germa89]

I presume we need to do nothing on the user side. I guess MAPDL is automatically switching to SMP? Or what happens if we have launched MAPDL with DMP mode and we want to run the krylov macros? I dont think we would get an error, would we?

[kmkoshy]

Hi @germa89 ,
If we launch the krylov macro in DMP, the user would see an error. I do not think it will fall back to SMP.
I think the only way is to force launch in smp mode using additional switches = '-smp' as suggested before.

[germa89]

Thank you @kmkoshy !

[germa89]

@FredAns @kmkoshy @dhabar95 please see:

#2326 (comment)

[ming0109]

I recommend waiting for the FULL file to be generated

mapdl.run('/SOLU')
mapdl.antype('HARMIC')  # Set options for harmonic analysis
mapdl.hropt('KRYLOV')
mapdl.eqslv('SPARSE')
mapdl.harfrq(0,1000)    # Set beginning and ending frequency
mapdl.nsubst(100)       # Set the number of frequency increments
mapdl.wrfull(1)         # Generate FULL file and stop

# Wait for the FULL file to be generated
import time
time.sleep(5)

mapdl.solve()
mapdl.finish()

dd = mapdl.krylov       # Initialize Krylov class object

[germa89]

In theory, it should not make a difference I guess.... did you find out that initializing the class later improve performance?