[Mesa-users] Questions about mesa-docker

[Evan Bauer]

Hi Ian,

I don't happen to know in particular how the threading interacts with pgstar windows, but I think the number of windows should have negligible impact on runtime. There does seem to be some non-negligible overhead associated with connecting to the X Server for pgstar windows, but I don't think 1 vs N windows makes much difference.

The main advantages of OpenMP come from parallelizing the most compute-intensive parts of the stellar evolution calculation, including things like EOS and opacity calls over all zones in the model, and the Jacobian matrix solves for the Newton iterations. You can get some sense of where in the code things are parallelized by looking for loops that are wrapped with OMP parallel instructions:
	grep "OMP PARALLEL DO" $MESA_DIR/star/private/*

Cheers,
Evan

> On Nov 8, 2024, at 5:07 PM, Ian Foley <ifoley2008 at gmail.com> wrote:
> 
> Hi Evan,
> 
> Thanks for this response. Hyper_V allows me to allocate up to 22 threads, but there is no setting for cores. When I run MESA, I have to set OMP_NUM_THREADS=n. With n set to 22 it ran much slower than setting it on 12. It seems to me that Hyper_V sets the maximum number of threads that can be allocated, but the actual number used depends on the OMP_NUM_THREADS set for the run. I don't know how many cores I am allocating to docker containers, only the number of threads. I'm guessing that 14 is the optimal number for me because there are 6 performance cores and 8 efficient ones. But it is a guess and I understand little about the processor architecture.
> 
> I did think the advantage of more threads would depend on how the MESA code has been written to incorporate OpenMP and that you might know about that. 
> 
> Am I correct to assume each pgstar window would run in a different thread so the number of pgstar windows and using pgstar will have little impact on MESA execution speed if there are enough cores? Are there other components of a run which will execute in parallel threads? What features of the code can be executed in different threads and so improve speed of execution if there are enough cores?
> 
> I don't know how to change the default settings with wsl2, and it appears to be better to use Hyper_V. There is some discussion about setting global defaults in a wslconfig file, but the format for its content is unclear.
> 
> Kind regards
> Ian
> 
> 
> 
> On Sat, 9 Nov 2024 at 07:44, Evan Bauer <evan.bauer.astro at gmail.com <mailto:evan.bauer.astro at gmail.com>> wrote:
>> Hi Ian,
>> 
>> Thanks for running these tests! I don't currently have a windows 11 machine to test all this on, so I probably don't have more concrete guidance to offer on the optimal configuration.
>> 
>> My intuition (similar to yours) might have been that you would want to configure MESA to run on the same number of threads as you have performance cores, but it sounds like it isn't quite that simple. It may be the case that the OMP thread scheduling for MESA parallelization is smart enough to make use of the efficient cores without starving the more powerful cores of useful work to do. So perhaps the optimal number of threads for your setup is 14 or 16? (I don't think it would be 22 because hyperthreading shouldn't gain you much for MESA runs.)
>> 
>> Do you know how many total cores you are allocating to docker containers? Sometimes that defaults to a lower number than the actual hardware cores that your system has.
>> 
>> Cheers,
>> Evan
>> 
>> 
>>> On Nov 6, 2024, at 3:48 PM, Ian Foley <ifoley2008 at gmail.com <mailto:ifoley2008 at gmail.com>> wrote:
>>> 
>>> Hi Evan,
>>> 
>>> Main question. What is the resource configuration in a Windows 11 pro container that will give the best performance when running MESA? I think the answer will depend on how the MESA software has been written to take advantage of multi-core processors.
>>>  
>>> My experience, reflected by the tests below, shows that the resource configuration can have a big impact on performance.
>>>  
>>> My configuration is a laptop running MESA in a container set up by Evan Bauer on Windows 11 Pro
>>> 
>>> The 5 tests below raise questions about the best configuration
>>>  
>>> Run 1 – Laptop was using an i9 intel processor with 6 cores and 12 logical processors and 16 GB RAM. Used Mesa version 21.12.1 with OMP_NUM_THREADS=6
>>>  
>>> Run 2-5 on a laptop using the intel ultra 9 185H processor with 32 GB RAM
>>> Processor Specifications:
>>> Total Cores 16
>>> # of Performance-cores 6
>>> # of Efficient-cores 8
>>> # of Low Power Efficient-cores 2
>>> Total Threads 22
>>> Max Turbo Frequency 5.1 GHz
>>> Intel® Turbo Boost Max Technology 3.0 Frequency 5.1 GHz
>>> Performance-core Max Turbo Frequency 5.1 GHz
>>> Efficient-core Max Turbo Frequency 3.8 GHz
>>> Low Power Efficient-core Max Turbo Frequency 2.5 GHz
>>>  
>>> Performance tests were made running test suite 1M_pre_ms_to_wd.
>>> I computed the times to run the 6 inlist’s in the test suite
>>>  
>>> Run 2. Set up container to use WSL2 as this was recommended by the Docker Engine. I found it was quite difficult to configure the resources. Roughly, the performance was twice as slow as that in Run 1 with the old laptop. Really surprised as I expected a significant improvement with the more powerful new laptop.
>>>  
>>> Run 3. Thought to setup container with Hyper-V. Found that it was easy to configure the resoueces. Set it up with 22 threads, 15.7 GB RAM and 3GB of swap space. Only did a partial run but found that the performance was about the same as the old laptop. Again not the improvement I expected. I wondered whether if I reduced the max number of threads it would be faster.
>>>  
>>> Run 4. Complete run with max of 12 threads. It ran twice as fast! Since my new laptop has a processor with 6 performance cores, I wondered if setting OMP_NUM_THREADS=6 will perform even better.
>>>  
>>> Run 5. Max of 6 threads.
>>>  
>>> inlists
>>> Run 1
>>> Run 2
>>> Run 3
>>> Run 4
>>> Run 5
>>> Mesa Version
>>> 21.12.1
>>> 24.3.1
>>> 24.3.1
>>> 24.3.1
>>> 24.3.1
>>> Configuration
>>> Hyper_V
>>> Wsl2
>>> Hyper_V
>>> Hyper_V
>>> Hyper_V
>>> No of threads
>>> 6
>>> 22
>>> 22
>>> 12
>>> 6
>>> Time unit
>>> mins
>>> mins
>>> mins
>>> mins
>>> mins
>>> Start
>>> 0.02
>>> 0.04
>>> 0.02
>>> 0.02
>>> 0.02
>>> End h burn
>>> 5.67
>>> 14.48
>>> About 5
>>> 3.92
>>> 6.19
>>> End he burn
>>> 446.14
>>> 525.24
>>>  
>>> 213.67
>>> 317.75
>>> He core flash
>>> 117.33
>>> 70.09
>>>  
>>> 34.93
>>> 47.47
>>> End agb
>>> 29.36
>>> 19.35
>>>  
>>> 10.72
>>> 14.52
>>> To wd
>>> 11.34
>>> 13.24
>>>  
>>> 4.77
>>> 6.36
>>> Total time
>>> 609.86
>>> 642.44
>>> ?
>>> 268.03
>>> 392.31
>>>  
>>>  
>>> Overall then using Hyper_V Manager seems to be 2 to 3 times faster than wsl2.
>>>  
>>> Also, with the processor I have, around 12 threads was better than 22 or 6. The reason for this is unclear and advice would be useful.
>>> 
>>> btw, I don't think the latest release has yet been added to mesa-docker.
>>> 
>>> Kind regards
>>> Ian
>>> 
>> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.mesastar.org/pipermail/mesa-users/attachments/20241111/d85dc45f/attachment.htm>