[Mesa-users] Same models on a recent MESA release runs much slower than on an old one

[Francis Timmes]

hi lorenzo,

please keep all replies on mesa-users. i got a chance to look and run.

these are bare-bones inlists, for both 11701 and 24.03.1, that go from creating an initial 1.7 msun model through core h burning, building of the he core through thin shell h-burning, the he-flashes, core he depletion, and finally to the first thermal pulse where they terminate. one can expect such a run to take some time, especially through the more challenging shell h-burning and core he ignition phases. this is what i saw when i ran the bare-bones 24.03.1 inlist.

if you want to reduce the run time, then you'll need to experiment beyond using pure default inlists. one way to do this is by watching the terminal output on what is limiting the timestep during which phase of evolution. for example, during the building of the he core through thin shell h-burning the terminal output says "dX/X h1" is limiting the time step. exploring the timestep controls

https://docs.mesastar.org/en/release-r24.03.1/reference/controls.html#timestep-controls

one find a "dX/X" timestep limiter corresponds to "dX_div_X_limit". searching on this latter term leads to the group of abundance timestep conrols 

https://docs.mesastar.org/en/release-r24.03.1/reference/controls.html#dx-limit-species

and then specifically 

https://docs.mesastar.org/en/release-r24.03.1/reference/controls.html#dx-div-x-limit-min-x

https://docs.mesastar.org/en/release-r24.03.1/reference/controls.html#dx-div-x-limit

note that sometimes the total run time can decrease when making the tolerances stricter. while more model steps are thus needed with stricter tolerances, each step converges quicker. this is can be more efficient for run time purposes than attempting larger timesteps where the solver may then struggle mightily to find a converged solution - which happens several times with the 24.03.1 bare-bones inlist approach to this 1.7 msun model.

happy experimenting!

fxt








> On Mar 27, 2024, at 5:51 PM, Lorenzo Martinelli <Lorenzo.Martinelli at uon.edu.au> wrote:
> 
> Hi Francis, 
> 
> Thank you for your answer. 
> I deactivated gold tolerances in both cases by setting use_gold_tolerances = .false. (in facts they were not used in the old models that I’m trying to port to the latest release) in order to check if that was the source of the difference, but the speed difference is still there.  
> Is there something else that I should try to tweak to find the source of the longer runtime?
> 
> I attached the inlists with use_gold_tolerances = .false. that still have a very different runtime starting from the RGB.
> Thank you,
> 
> Lorenzo Martinelli
> 
> 
>> On 28 Mar 2024, at 04:03, Francis Timmes <fxt44 at mac.com> wrote:
>> 
>> hi lorenzo,
>> 
>> 11701 was released on 03may2019. since then, the default convergence tolerances that mesa uses have gotten significantly tighter (e.g., the gold and gold2 tolerances, https://docs.mesastar.org/en/release-r24.03.1/reference/controls.html#use-gold-tolerances). the cost of these more converged solutions can be longer runtimes. one can, of course, change these convergence tolerances ...
>> 
>> fxt
>> 
>> 
>> 
>> 
>> > On Mar 27, 2024, at 5:55 AM, Lorenzo Martinelli via Mesa-users <mesa-users at lists.mesastar.org> wrote:
>> > 
>> > Dear MESA Users, 
>> > 
>> > I’m porting to the recent MESA r23.05.1 (and lately r24.03.1) some of the inputs and custom subroutines in the run_star_extras used for computing large grids of models with an old release MESA r11701.
>> > I’ve been successful in porting all the features, the models run, behave and and look as expected. However I noticed that especially for models that develop strongly degenerate helium cores after the main sequence, roughly below 2.0 solar masses, the new models were running significantly slower.
>> > After making sure that none of the custom subroutines and inputs were the cause of this slowness of the models in the latest version, I copied the default /mesa-r23.05.1/star/work and just change the initial mass to 1.7 solar masses. I did the same for MESA r11701, run both of them and compared the runtime.
>> > 
>> > I found that even with all default settings, the most recent one was running much slower (on the MS it is not very noticeable, but it starts to be evident on the RGB), with a difference of several hours (a big deal when you have to run a grid). 
>> > 
>> > I attach here the inlists that I used for these tests, which I just replaced inside copies of the /mesa-r23.05.1/star/work and /mesa-r11701/star/work respectively.
>> > 
>> > I’m using the appropriate MESA SDK for each different MESA release and I’m running them on the same MacBook Pro Quad-Core Intel Core i7, I even tried to run the MESA r23.05.1 also on a new MacBook with M3 chip (8 cores) and it’s still slower than the old MESA r11701 on the Quad-Core Intel Core i7.
>> > 
>> > Does someone have an idea why the models with the more recent MESA release runs slowe r even with the defaults settings? Is there something that I can do to recover that speed (which is quite critical considered the number of models that I need to compute)?
>> > Thank you!
>> > 
>> > Kind regards,
>> > 
>> > Lorenzo Martinelli
>> > <inlist_project_r23.05.1><inlist_project_r11701>_______________________________________________
>> > mesa-users at lists.mesastar.org
>> > https://lists.mesastar.org/mailman/listinfo/mesa-users
>> >
> 
> <inlist_project_11701><inlist_project_r23.05.1>