How to run parametric analysis in parallel?

[yang-sjtu]

Dear all,

In previous single analysis, it works fine to run the simulation on 4 cores in parallel by setting the cores= tasks=4 ; however in the Parametric analysis, set cores=4 tasks=4 leads to running 4 tasks of parametric sweep simultaneously.

Moreover, set cores = 6 and tasks = 2, leads to running 2 tasks of parametric sweep at the same time, but each task runs in serial, namely only 2 cores are active, while 6 activated cores and 3 cores for each task are expected.

variable_dict = ipk.parametrics.add_from_file("Test_Table.csv", name="Parametric_file")

ipk.analyze(setup="Parametric_file", cores=6, tasks=2)

Anyone knows how to setup parametric analysis so that the tasks can be solved in parallel properly?
For example, using 6 cores to run 2 tasks at a time and 3 cores for each task in parallel.
Many thanks!

[yang-sjtu]

I notice there is a proposal similar to my purpose Add more options to submit_job() #6316
.

I'm wandering how could this be implemented at this moment, is there any way to work around? Any help would be great appreciate!

[JonathonMcMullan]

As a short term fix if it isn't respecting your ACF configuration, you could initialise the older analyze functions within PyAEDT. It didn't copy very well in the original post but I was modifying/executing the ACF config with the older commands while doing everything else with native PyAEDT commands. So I would use the method above with whatever config you needed (including the "NumVariationsToDistribute" parameter). You would get the "Optimetrics" module (oDesign.GetModule("Optimetrics")), then execute the parametric study you setup with oModule.SolveSetup("ParametricSetup1")

[yang-sjtu]

Thanks! Apparently, I don't know how to implement the parallel execution of subtasks in parametric analysis in GUI as well. For example, whether or what to set for "NumVariationsToDistribute", "Enable2LevelDistribution" or "NumL1Engines" in GUI.

If that first step succeed, I assume your solution can be adopted to convert it into automation code in script/python, with all the necessary parameters passed properly. Thanks again!

Indeed, since I implement this work directly on PyAEDT, it pops up like what I post in the beginning.

[JonathonMcMullan]

2TaskAuto.txt
2TaskManual.txt

As a test case you could try something like the following. Do all of the code you need to do between the "ansys.aedt.core.Hfss()" line and the "hfss.odesktop" line including a solve setup. Then you will need to complete the "c.InsertSetup('OptiParametric')" line based on the formatting in the HFSS scripting guide (or you can make this with the PyAEDT method). Then the "c.SolveSetup" line will execute the sweep with 2 tasks in parallel with 4 cores each based on the acf file specified in the acf_path variable.

I haven't checked it since I worked on a benchmarking project, but its worth checking if the standard PyAEDT methods have resolved the issue the example below was originally created to fix.

import ansys.aedt.core
import os

hfss = ansys.aedt.core.Hfss()

a = hfss.odesktop
b = a.odesign

acf_path = r"C:\Users\Jonathon.McMullan\Desktop\2TaskAuto.acf"

a = hfss.odesktop
b = hfss.odesign
print(a.SetRegistryFromFile(acf_path))
print(a.SetRegistryString('Desktop/ActiveDSOConfigurations/HFSS', '2TaskAuto'))
print(a.GetRegistryString('Desktop/ActiveDSOConfigurations/HFSS'))
c = b.GetModule('Optimetrics')
#c.InsertSetup('OptiParametric')
#c.SolveSetup('ParametricSetup1')

hfss.close_desktop()

[yang-sjtu]

Hi Jonathon,

thank you so much for your replies and inspiring discussion. Following your suggestion yesterday, I tried to work with ACF file out and finally find the solution after some demo retry.

One more thing I have to clarify is what I work on is Icepak, I should have mention it more clearly in the very beginning. And the different solvers behind Icepak (FVM solver from Fluent) and HFSS (FEM solver from Ansys) make our discussion a little bit less straightforward :)

In Icepak, SMP with MPI calls No. of processes = No. Tasks in Setup to run a single simulation case in parallel. Therefore, the No. Tasks = No. Cores is still necessary, and my purpose in Icepak could be achieved by modified the following parameters in the acf_file I post above:

...
					NumEngines=6
					NumCores=6
...
                                AllowedDistributionTypes[3: 'Variations', 'Mesher', 'Solver']
				Enable2LevelDistribution=True
				NumL1Engines=2
···

On the contrary, in HFSS it might be case to use No. Tasks=1 for a single simulation case to run on No. Cores in parallel, then I guess your acf_file demo can be used for a similar purpose in HFSS.

Your help is really appreciated!