[Mesa-users] [EXT] Question about scaling of the PPISN test suite

Thu Sep 12 11:38:01 UTC 2024

Hi Jeremy,

Mesa TestHub results (for ppsin) show there is a speedup going from 8 to 16
cores. From about 90 minutes to 60 min. So the scaling improvement is 1.5,
not 2, for doubling the number of cores, due to constant overhead that
cannot be parallelized away. The details depend on the machine, of course,
so you'd have to check with your machine of course.

I might give 16 or 12 cores a try over 19 if it means you can fit more jobs
on a computer node.

Ref.

https://testhub.mesastar.org/main/test_cases/star/ppisn?utf8=%E2%9C%93&history_type=show_instances&start_date=2023-07-01&end_date=2024-09-13&computers=cannon_conroy_intel&status=&commit=Search

(Make sure the 'threads' checkbox is checked, and compare the 'full' tests
o ly with the little green plus signs)

Best,
Philip

On Wed, Sep 11, 2024, 9:15 PM Renzo, Mathieu - (mrenzo) via Mesa-users <
mesa-users at lists.mesastar.org> wrote:

> Order of 10s of hours (it also depends how many pulses a model
> encounters).
>
> The memory requirements will depend on your spatial resolution and the
> nuclear reaction network mostly.  10GB should be sufficient unless you use
> large nuclear networks and/or very high resolution. In my experience the
> cpu is typically more stringent requirement than the memory.
>
> From tiny keyboard
>
> ------------------------------
> *From:* Jeremy Sakstein <sakstein at hawaii.edu>
> *Sent:* Wednesday, September 11, 2024 6:09 PM
> *To:* Renzo, Mathieu - (mrenzo) <mrenzo at arizona.edu>
> *Cc:* Jeremy Sakstein via Mesa-users <mesa-users at lists.mesastar.org>
> *Subject:* Re: [EXT] [Mesa-users] Question about scaling of the PPISN
> test suite
>
>
>
> * External Email *
> ------------------------------
> Thanks Mathieu,
>
> This is very helpful.
>
> Do you happen to recall how long a run took on 8 cores?
>
> Also, I meant to ask above: do you have any recommendations for the memory
> for PPISN? I usually request 10GB but this is again not based on detailed
> experiments.
>
> Cheers,
>
> Jeremy
>
> On Wed, Sep 11, 2024 at 3:05 PM Renzo, Mathieu - (mrenzo) <
> mrenzo at arizona.edu> wrote:
>
> Hi Jeremy,
>
> The maximum efficiency will likely be sensitive to your specific setup and
> MESA version, and the experiment I did with MESA 11123 are probably not
> informative anymore/at all. For what it's worth, i ran most of my grids
> back then using 8 cores, so that I could fill nodes of the cluster with an
> integer number of models (and threw away many failing models).
>
> It also depends on the size and dimensions (as in number of parameters you
> vary) in your grid: the rule of thumb is that for a given number of
> available cores, fewer cores per model to run more models in parallel gets
> you the grid faster (see for example Rob Farmer's examples setting
> OMP_NUM_THREADS=1 https://cococubed.com/mesa_market/many_jobs.html
> <https://urldefense.com/v3/__https://cococubed.com/mesa_market/many_jobs.html__;!!PvDODwlR4mBZyAb0!VkkzZKyZJvimPltkyP-lGErTtUlxkRslra616PYal-kcdXZ14nvuyM3pI_PAKQAbH2WPjKR5fRW9UUPiAOo$>
> ).
>
> All that being said, 19 cores is probably a lot (and a number that will
> probably not optimize the number of models you can fit on each node of a
> cluster). But you'll have to profile your setup yourself and consider the
> number of core per node available at the computing facility you have access
> to.
>
> Hope this helps!
> Mathieu
>
> From tiny keyboard
>
> ------------------------------
> *From:* Mesa-users <mesa-users-bounces at lists.mesastar.org> on behalf of
> Jeremy Sakstein via Mesa-users <mesa-users at lists.mesastar.org>
> *Sent:* Wednesday, September 11, 2024 5:35 PM
> *To:* mesa-users <mesa-users at lists.mesastar.org>
> *Subject:* [EXT] [Mesa-users] Question about scaling of the PPISN test
> suite
>
>
>
> * External Email *
> ------------------------------
> Hi everyone,
>
> I am planning to run a large grid of models using a work directory based
> on the PPISN test suite.
>
> Has anyone experimented with the efficiency of parallelization for this? I
> typically run on 19 threads but this is not based on anything quantitative.
>
> I'd like to make the runtime as efficient as possible.
>
> Cheers,
>
> Jeremy
>
> --
> Jeremy Sakstein (He/Him/His)
> Assistant Professor of Theoretical Physics
> Department of Physics & Astronomy
>
> <https://www.google.com/maps/search/2505+Correa+Road+%0D%0A+Honolulu,+HI,+96822,+USA?entry=gmail&source=g>University
> of Hawaiʻi at Mānoa
> Watanabe Hall
> 2505 Correa Road
> <https://www.google.com/maps/search/2505+Correa+Road+%0D%0A+Honolulu,+HI,+96822,+USA?entry=gmail&source=g>
> Honolulu, HI, 96822, USA
> <https://www.google.com/maps/search/2505+Correa+Road+%0D%0A+Honolulu,+HI,+96822,+USA?entry=gmail&source=g>
> http://www.jeremysakstein.com
> <https://urldefense.com/v3/__http://www.jeremysakstein.com/__;!!PvDODwlR4mBZyAb0!VkkzZKyZJvimPltkyP-lGErTtUlxkRslra616PYal-kcdXZ14nvuyM3pI_PAKQAbH2WPjKR5fRW9x89RKeY$>
>
>
>
> --
> Jeremy Sakstein (He/Him/His)
> Assistant Professor of Theoretical Physics
> Department of Physics & Astronomy
>
> <https://www.google.com/maps/search/2505+Correa+Road+%0D%0A+Honolulu,+HI,+96822,+USA?entry=gmail&source=g>University
> of Hawaiʻi at Mānoa
> Watanabe Hall
> 2505 Correa Road
> <https://www.google.com/maps/search/2505+Correa+Road+%0D%0A+Honolulu,+HI,+96822,+USA?entry=gmail&source=g>
> Honolulu, HI, 96822, USA
> <https://www.google.com/maps/search/2505+Correa+Road+%0D%0A+Honolulu,+HI,+96822,+USA?entry=gmail&source=g>
> http://www.jeremysakstein.com
> _______________________________________________
> mesa-users at lists.mesastar.org
> https://lists.mesastar.org/mailman/listinfo/mesa-users
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.mesastar.org/pipermail/mesa-users/attachments/20240912/d3a19f87/attachment.htm>