[toboto]

Dear All,

Is there any benchmarks for the scalability of Flow3d v11.2 (NOT MP version) on clusters.

I am asking since I got very disappointing results myself. I tried identical model on two machines:

• 1st (laptop) with Intel Core i7 8th Gen (4 cores - 8 threads) + 16 GB ram
  
• 2nd (Workstation) with dual Intel Xeon E5-2690 V3 (24 cores - 48 threads) + 32 GB ram

I only gained about 10% acceleration with the workstation.

Is that normal with the non-MP version of Flow3d?

Is there any way to improve the performance on the workstation without the need to buy Flow3d/MP license?

Best regards

[M.Nazrin]

Hi Toboto,

For workstation, how many ram slot does the motherboard have and how many are populated with ram?

[toboto]

Originally Posted by M.Nazrin

Hi Toboto,

For workstation, how many ram slot does the motherboard have and how many are populated with ram?

There are 4 ram slots for each processor. So the workstation have 8 ram slots. Currently it has 4 slots (2 per processor) occupied with 4x 8 GB ram.

[flotus1]

I guess the non-MP version still runs in parallel, just with OpenMP instead of a hybrid OpenMP/MPI parallelization. So my advice will be based on that assumption.
In this case, you could first try to pin the simulation to one of the two CPUs. This will prevent high memory latency from from NUMA. How to do this depends on the operating system. "taskset" is commonly used in linux. And of course, only use 12 threads for this simulation. Additionally, disable hyperthreading in the bios, which is a good idea for CFD workstations anyway.
If this does not speed things up, you can try to play around with core binding options, to prevent the OS from shuffling around threads to different cores every few seconds. Contact support to check how it could be done.
If that still doesn't work, your case might just be too small to scale on many cores. Or so large that both machines run out of memory.
Edit: or some other bottleneck, like file I/O. Or your time measurement might include a lot of single-threaded setup time...

[Habib-CFD]

Flow 3D software in non-HPC version 11.2 is somehow optimized for 8 cores. The performance acceleration falls drastically after 8 cores even using single EPYC 7302P occupied by 8 channel memory.

This is off topic but for your information and comparison: I had a very nice experience using 2920x in 8 cores, frequency limited to 3.8 GHz, vcore 1.08v, 4x8gx3200MHz of RAM. The simulation time was about 15 percent slower than EPYC 7302p by 16 cores.

In my opinion, a new PC like an Intel 9800X light workstation will provide much more performance than paying for the HPC license in your old machine (with lower memory frequency) unless you plan for a large model.