[Stabum]

Hello everybody,
I've seen a very interesting video on YT, in which a CFD simulation was launched in Fluent 18.2 using an increasing number of cores. The simulation (always the same in terms of settings) was faster as the core number had been increased (2 cores, 3 cores, 4 cores). After the 4th, with 5, 6, and more cores, the simulation time was always the same and did not show any improvement. Then the guy increased the amount of RAM used and relaunched the simulations all over again. And he descovered that the simulation times, this time, continued to decrease even with 7, and 8 and more cores used.
At the end of the day, it seems that, if you have 4 or more GB of RAM per core used, the simulation times continue to decrease: the added core "works" and the curve "sim time vs number of cores" has a negative slope.

Now, let's imagine to use a Threadripper 3970X (32 cores) CPU. If we want to make all these cores effective and useful in terms of simulation times reduction, we should then have 32x4=128 GB of RAM.

Have you ever faced this RAM/core correlation?
I'm really curious of reading about your personal experiences.

Thanks a lot,
kind regards!

[flotus1]

I'm afraid that's not how things work with Fluent, or any other CFD solver. Memory per core is pretty much a meaningless metric. Either the simulation fits into RAM, or it doesn't.

That guy on youtube with undisclosed hardware most likely ran into one of two issues, or a combination of both:
1) His simulation didn't fit into memory before upgrading it. This is rather unlikely, because the performance impact would be so severe that it should raise questions about the validity of the test
2) By upgrading to more memory, he accidentally unlocked additional memory channels. This can actually speed up CFD simulations, and enable scaling up to higher core counts.

Edit: I think I found the video you are referring to. It is packed with poor practices and misleading information. My brutally honest opinion: try to forget what you took away from this particular source of information.

[Stabum]

Originally Posted by flotus1

I'm afraid that's not how things work with Fluent, or any other CFD solver. Memory per core is pretty much a meaningless metric. Either the simulation fits into RAM, or it doesn't.

That guy on youtube with undisclosed hardware most likely ran into one of two issues, or a combination of both:
1) His simulation didn't fit into memory before upgrading it. This is rather unlikely, because the performance impact would be so severe that it should raise questions about the validity of the test
2) By upgrading to more memory, he accidentally unlocked additional memory channels. This can actually speed up CFD simulations, and enable scaling up to higher core counts.

Hello, and thanks a lot for your answer.
I understand point (2) of your answer. I don't understand the first one. I mean: if it's true what you say (Either the simulation fits into RAM or not) and I agree, the point is that the simulations (no matter how many cores were used) did start and came to the end. As long as the number of cores used was 4, every time a core was added, the simulations were faster; with 5 or more cores, the simulation time remained the same.

Thanks again


ref. https://www.youtube.com/watch?v=NguWpYd5daY
time: 3:18

Please note that in the two graphs the duration of the simulation with 1, 2, 3, 4 CPUS is exactly the same.