Threadripper Optimization

[Dario Proietti]

Hello everyone,

i have some problem running test on my threadripper 5995wx machine (64 core,128 thread) and 2950x (16core, 32 thread) , it's very slow, i already have a ryzen 7950x (16core, 32Thread). 

With Ryzen compute time is like half from threadripper. 

i only use 1 mesh for simulation.

how can i optimize it?

running on windows,  i try to create a specific cpiinfo file but nothing change

[Randy McDermott]

For only one mesh, you should use up to 8 openmp threads.  Look at the user guide for how to launch a job using openmp.  You won't see more than a factor of 2 speed up.

If you have a 64 core machine, your best speed up would be to use these cores for difference meshes (with different MPI processes).  This scales well until you have about 5000-10000 cells on a mesh.

--
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 view this discussion on the web visit https://groups.google.com/d/msgid/fds-smv/cf4f7416-2e76-4282-a3f2-873e86a7e674n%40googlegroups.com.

[Dario Proietti]

its important to have a speedeer ghz cpu or not? this is the reason because ryzen is faster then threadripper cpu in single mesh?

if i wanna buy a new machine which is the better configuration (single mesh)?

[Randy McDermott]

For a single mesh (big constraint), you do not get much speedup beyond 4 openmp threads (1 thread per physical core)---this maxes out at a factor of 2 speedup.  Most cpus have many more cores than that these days.  So, a small number of cores with fast clock rate and fast internal memory access is what you want to target.

But you should be aware that you will get much better performance by utilizing the cores as MPI processes, breaking up the domain into multiple meshes.  In that case, if you have a cpu with, say, 16 cores, as long as you have enough RAM, you can run a case with 16 meshes with at least 1000 cells/mesh and get much closer to linear speedup, meaning 16 times faster than a single mesh with one openmp thread.

To view this discussion on the web visit https://groups.google.com/d/msgid/fds-smv/1fefb24f-c5b5-4ba8-8e6b-9afe235ea96bn%40googlegroups.com.

[Dario Proietti]

I use only one mesh because I realized that when I create automations, for example if in the event of a fire on the ground floor some doors on the upper floor have to lock, if these are not in the same mesh they do not communicate with each other. I need to solve this problem otherwise I'm forced to always work in single mesh, make the meshes communicate with each other. it's possible?

[Randy McDermott]

What version of FDS are you using?  If the latest (6.8.0), then submit a simple case to the Issues page.

To view this discussion on the web visit https://groups.google.com/d/msgid/fds-smv/1258e13e-d73f-4458-bcdf-529cf94eb300n%40googlegroups.com.

[Dario Proietti]

no I'm using older version (6.7.5 or 6.7.0) because the bim program I use does not support the latest versions