[HyperNova]

Hello experts,
currently i am working with a computer which has 2*intel Xeon 4669v3 (each cpu has 18 cores and 36 threads) with 512Gig of RAM.
I set up 3 fluent works together and i give 20 core each of them. but when i check the task manager it seems only half of the threads are computing the solution and those seem to be located in one of the two cpus. That cause the simulation goes very slow . Any suggestion would be appreciated. Windows server 2012 R2 is installed on the system if it helps.
Another important question: system has Tesla K40 GPU. I have installed the drivers and when i start the Fluent there is no massage in consul that GPU is not working. But when I start the simulation there is no sign that GPU is handling any part of simulation.
I would be grateful if experts in this forum or anybody who had also experience with the issues help me. sorry if there would be some very basic parts in my problem.
Regards

[CapSizer]

OK, so the best advice I can give to you is to build up an understanding of how well Fluent is scaling on your workstation. So take one fairly large problem, preferably start from a calculation that has run a few 100 iterations, then time how long it takes for another 100 iterations over 4 processes, then 8, then 12, etc., all the way to 72. Plot the number of iterations per hour versus the number of compute cores.

Ansys actually provides some standard benchmark cases that you can use, as well as -benchmark option which will run a range of numbers of compute processes. Otherwise just check with one of your own cases.

My guess is that you will see essentially no improvement from 36 to 72 processes, because hyperthreading is not an effective way of making Fluent run faster. On our systems we keep hyperthreading off completely.

I also expect diminishing returns over the range from about 16 to 36 processes. The reason for this is that the system simply cannot provide adequate memory bandwidth for all those processor cores. You should find pretty much linear improvement up to about 16 processes. I hope.

But do the investigation and please report on your findings.

If you want to use the GPU, you need to set Fluent up to use it when you launch the run. Once again, if you get it working, please report on your findings. I haven't had much luck with getting meaningful acceleration, but I was using a much older card.

[Micael]

Fluent does not scale on "thread", first thing is to disable hyperthreading. You will be left with 36 core and should see good scaling up to that number. Keep in mind that under Windows you should reserve at all time at least 1, preferably 2 core for operating system and stuff. So you basically really have 34 core for simulation on your system.