remove hyperthreading with MPI

[luca.mi...@gmail.com]

Hi all,

I am trying to understand the best setup for my new workstation  (CPU Intel(R) Xeon(R) W-2195 CPU @ 2.30 GHz, 32 GB) in order to give a speedup to my simulations.

I have found these very useful notes https://github.com/firemodels/fds/wiki/OpenMP-Notes . 

I didn't understand if I should remove hypertreading also if I am using MPI, or if it is better to remove hyperthreading only if I am using OpenMP. I have found some discussions in this forum about this topic, but I still have some doubts.

Thanks

[Glenn Forney]

Remove hyperthreading. It doesn't help openmp or openmpi. 

--
You received this message because you are subscribed to the Google Groups "FDS and Smokeview Discussions" group.
To unsubscribe from this group and stop receiving emails from it, send an email to fds-smv+u...@googlegroups.com.
To post to this group, send email to fds...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/fds-smv/00a4a34b-f71f-4fc0-b447-143fb0509602%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[luca.mi...@gmail.com]

Thank you Glenn for your answer.

So the better strategy in my case, if I have a big model, is to remove hyperthreading and , since my workstation has 18 core, I should divide the model in 18 meshes and use only MPI.

Otherwise if the model is not that big, I can add OpenMP parallelization.

Could this be a good strategy?

[Salah Benkorichi]

log into your bios and disable hyperthreading, it's not good for run time whether for small or big jobs.

MPI will always yield better than OpenMP.

Having 18 cores doesn't guarantee too much of speed, from my limited experience I would say having 12 cores machine is usually sufficient to be considered as one node, and usually this will come into two sockets each with 6 cores. Filling your sockets will actually not help improving the run time.

The best strategy that I have is through trial and error, which means that if I have large model, I would set it to around 2-5s only, and test different settings, nbr of MPI used, or number of nodes used, and also nbr of OpenMP used, generally when having large model, always set OMP_NUM_THREADS=1, whatever setting yield better than I choose it. You might lose around 20-30min setting it but will gain you hours if not days if the model is quite large.

To unsubscribe from this group and stop receiving emails from it, send an email to fds-smv+unsubscribe@googlegroups.com.

To post to this group, send email to fds...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/fds-smv/6aa9d776-3f81-4b63-a752-a579c9527829%40googlegroups.com.

[luca.mi...@gmail.com]

Thank you very much Salah for the answer and for sharing your experience.