Number of processors makes a big difference!

I was solving something this morning and had previously done a benchmark that I found here in the forum not long ago. The benchmark indicated that my computer’s ideal number of cores was 16. I’ve been using that number for a while not really understanding what the baseline meant. But today I increased the number to 46, then I stopped the analysis and put it back to 16. I was thinking that because the processing graph at 16 never got to the 100% that it was really not ideal but check it out:

image

Each tiny bump on the right using 16 cores equals the same calculation on the left side using 46 cores. The same for my desktop which only has 16 total cores, the benchmark showed that 8 was ideal. So its it very important to know and use this number you guys!.. its probably just me, I learn things the hard way.

1 Like

Here’s a nice benchmark (description in Polish): CalculiX solver

It’s usually advised to use physical cores for parallelization but I always just enter the number of logical CPUs when running FEA.

1 Like

This benchmark is highly misleading and inaccurate. It fails to distinguish between physical and virtual cores. While I’m not aware of imgprojts PC’s configuration, I suspect his observation may be related to memory bandwidth saturation. Therefore, depending on the hardware, once a certain number of cores is reached, each increase in core count will asymptotically approach zero improvement in solve time.

1 Like

Yes that is absolutely true, FEM is very cache dependent in nature as well, more so than FVM codes and performance absolutely tanks when memory bandwidth is realized, or core saturation per channel