[Mesa-users] Optimizing MESA for HPC server with 128 cores and 128Gb memory

[Francis Timmes]

(n_isotopes + 5) * n_cells * (3*(n_isotopes + 5) + 9) < 2^31

for n_isotopes = 25, n_cells < 1e6

for n_isotopes = 45, n_cells < 270,000

for n_isotopes = 200, n_cells < 16,7878

an approximate formula is

n_cells < 7e8 / (niso + 5)^2 

fxt



> On May 16, 2023, at 6:16 AM, Levi via Mesa-users <mesa-users at lists.mesastar.org> wrote:
> 
> Dear Rob and Earl,
> 
> Thank you both for your helpful insights. Your guidance is greatly appreciated.
> 
> Best,
> Levi
> 
> ---- Replied Message ---- 
> From Earl Bellinger<earlbellinger at gmail.com> Date 5/16/2023 20:55 To Rob Farmer<robert.j.farmer37 at gmail.com> Cc Levi<lileiixx at 163.com> ,
> mesa-users at lists.mesastar.org<mesa-users at lists.mesastar.org> Subject Re: [Mesa-users] Optimizing MESA for HPC server with 128 cores and 128Gb memory Dear Levi,
> 
> Just to add a bit on Rob's answer:  
> 
> Depending on the use case, a good number of threads to use with MESA seems to be in the ballpark of around ~12. See Figure 47 in the paper Rob linked. 
> 
> With 128 threads it's possible you could even see a degradation of performance, because creating and managing threads requires computing resources. If you create too many threads, the overhead of managing them can outweigh the performance benefits of parallelism. If you wish, you could test the performance as a function of the number of threads like in the MESA instrument paper, and report back to us with the results. 
> 
> Note that if your ultimate goal is to construct a grid of >=128 tracks, then it will actually be fastest on your system to run each track with 1 thread. See e.g. Amdahl's law for more info. 
> 
> Best wishes, 
> Earl 
> 
> ---
> Earl Patrick Bellinger 
> Postdoctoral Research Fellow 
> Max Planck Institute for Astrophysics 
> https://earlbellinger.com 
> 
> 
> On Tue, May 16, 2023 at 11:36 AM Rob Farmer via Mesa-users <mesa-users at lists.mesastar.org> wrote:
> Hi Levi,
> 
> First off you can't use mesa3335.net in a mesa model. Limitations in how mesa handles arrays means the max number of isotopes is ~500 (depending on the number of mesh points).
> 
> If your machine has 128 cores then you just do:
> export OMP_NUM_THREADS=128
> 
> See Section 6 of https://ui.adsabs.harvard.edu/abs/2019ApJS..243...10P/abstract for a discussion of MESA's scaling with core counts. Short answer is that it is not linear, so don't expect MESA to use all available CPUs all the time.
> 
> Rob
> 
> 
> On Tue, 16 May 2023 at 10:41, Levi via Mesa-users <mesa-users at lists.mesastar.org> wrote:
> Dear  MESA-users, 
> 
> I have a question about using MESA on high-performance computing systems. I want to use the large reaction network mesa3335.net to evolve a model, but my local computer shows insufficient memory. So we switched to a server with 128 cores and 128Gb of RAM, but it still doesn't work, MESA does not seem to fully utilize the CPU and memory. 
> 
> How do we properly configure MESA to leverage the computing resources of this server? Specifically, how can we set up the parameters to ensure MESA calls on the multiple cores and large memory available? 
> 
> Any guidance you can provide would be greatly appreciated. Thank you in advance for your time and assistance! 
> 
> Best,
> Levi
> _______________________________________________
> mesa-users at lists.mesastar.org
> https://lists.mesastar.org/mailman/listinfo/mesa-users
> 
> _______________________________________________
> mesa-users at lists.mesastar.org
> https://lists.mesastar.org/mailman/listinfo/mesa-users
> 
> _______________________________________________
> mesa-users at lists.mesastar.org
> https://lists.mesastar.org/mailman/listinfo/mesa-users
>