Calculation of bremsstrahlung

[jonmaddock]

In GitLab by @mkovari on Jan 14, 2021, 14:43

Issue #1230 showed that function pbremden (Bremsstrahlung radiation density) takes 32% of the total execution time (includes calls from that function).

As far as I can see, the only purpose of this function is to be able to divide the total radiation into line radiation and bremsstrahlung so as to be able to output these quantities. I think that only the total radiation is actually used in the remainder of the PROCESS calculation. The only line I can find where the bremsstrahlung is used is

pline = pimp - pbrem

where

!! pimp  : output real : total impurity radiation density (W/m3)
!! pbrem : output real : Bremsstrahlung radiation density (W/m3)
!! pline : output real : other radiation density (W/m3)

If this is correct, then we have some simple options for speeding up the code:
EITHER Remove this routine entirely and remove the separate output of line radiation and bremsstrahlung,
OR calculate the bremsstrahlung only on the function call that occurs after the last VMCON iteration when the OUT.DAT will be written.

Detailed options:

1. Replace the function zav_of_te by something much quicker.
2. Reduce the number of steps in the profile integration: npts = 200 in subroutine imprad.

[jonmaddock]

In GitLab by @wg7580 on Jul 20, 2022, 15:19

Having profiled the code on the branch of the impurity_radiation.f90 conversion, using baseline 2018 and 2019, it does not seem as though this is an issue anymore.

[jonmaddock]

In GitLab by @wg7580 on Jul 20, 2022, 15:19

closed

[jonmaddock]

In GitLab by @mkovari on Jul 21, 2022, 13:18

Hi Sam, @wg7580 - can you upload the data showing that this is no longer an issue as far as performance is concerned? It seems implausible that converting to Python would cause this routine to be called less often. @timothy-nunn

Also - there is a lot to be said for removing the bremsstrahlung calculation anyway, even if it isn't causing a performance problem.

[jonmaddock]

In GitLab by @wg7580 on Jul 21, 2022, 14:16

Hi Michael- using a few input files with profiling it didn't show up to be using a lot of the run time. From the above image- it shows that the self time (in brackets) which is the time spent on the function itself is low- so must be spending time in zav_of_te as you said. I wondered if you had any idea which input file caused this to be so slow so i could try and replicate the error? When we have been profiling it has mainly been pf coil that is causing issues.

[jonmaddock]

In GitLab by @mkovari on Aug 18, 2022, 15:10

reopened

[mkovari]

I propose to remove the separate output of line radiation and bremsstrahlung entirely. Only the sum of the two and the synchrotron radiation will be reported. This is to save execution time. (If you read the comments above you may think this is not an issue, but I recall that further investigation showed that profiling carried out was incorrect in some cases.)

The line radiation and bremsstrahlung radiation are used to calculate hstar, the non-radiation corrected value of the energy confinement factor H. This only appears in the two-page summary, not in the OUT.DAT or the MFILE. Please let me know if you have any objection to my removing hstar.

@ajpearcey @jmorris-uk @ym1906

[mkovari]

@timothy-nunn Graeme has reminded me that this code was refactored when converting to Python. Are you able to run tools to assess whether the bremsstrahlung and impurity radiation calculations are occupying any significant execution time?

PS I am still minded to remove the bremsstrahlung code, since it is not needed and there is an issue suggesting it may be inaccurate (#283) and another one suggesting the calculation of hstar may be incorrect (#652).

[timothy-nunn]

@mkovari I will run this test again on Thursday. From memory this was no longer a large portion of our relative runtime. We can also maybe discuss the removal of it all together then :)?

[mkovari]

In large-tokamak,
imprad_profile is called 5456 times.
pbremden is called 5456 times.
zav_of_te_counter is called 54568 times.
zav_of_te_counter returns an array of size given by plasma_profile.profile_size = 501.
It has three lines of code with logarithms, so in all, I think,
3 x 501 x 5456 = 8,200,368 logarithms are calculated.

I don't think execution time is as big an issue as it was. However, I still propose to remove all calls to this routine and remove the separate output of line radiation and bremsstrahlung.

Note that the Bremsstrahlung radiation power is output for stellarators as well as tokamaks. Can someone confirm whether anyone wants or needs this output?
@ym1906 @ajpearcey

[grmtrkngtn]

I have no particular use for Bremsstrahlung as an output. If it is indeed inaccurate perhaps it is best to remove it.

[ajpearcey]

in PR !3110 which solves issue #438 I added the hstar calculation into the OUT.DAT

[ajpearcey]

Looking again at the hstar calculations that are now implemented in physics.py they only require the total core radiation and synchrotron radiation. So removing the Bremsstrahlung calculation will not effect it. Therefore if we think it best to remove the calculation it can easily be done.