[ettore]

Hi,
I have recently start using a 4P machine and for my surprise i came across a limitation (or a fault of mine in simulation build up). Although the machine has 64 cores and 128 GB of ram (4 memory channels per CPU, total of 16 that i was planning to use) i cannot spawn the simulation on more than 30 nodes. Anything above 30 nodes just ends up in a faulty cpu affinity set that messes up the simulation calculation time. While settingup 30 nodes the calculation runs as expexted (i had some benchmarks done in order to see the diffrence between various hardware).
Should it be a fault of mine an can anybody help me find out what is it that i am doing wrong?
I have researched the Fluent limitations but i couldn't find anything.
Should that be the max limit of nodes per machine are there any possibilities on manually picking the nodes? The ideea is to spawn the calculation on every CPU in order to profit of all 16 memory channels.
I should also mention that the OS is a Windows (that can see and use the 4P, 128 GB hardware). Other programs use / benefit of the complete hardware.
The simulation that i'm doing is an air flow in a closed chamber. Number of cells is about 21 mil.

I am very thankfull for any answeres and opinions.

[ettore]

After more research i found an aswere to my problem.
Fluent does spawn over more than 32 nodes but on a local machine the limit is 32. In order to have more than 32 active nodes you need at least another machine (or more).
In conclusion should you decide on a 4P machine go for 8 Core CPUs (resulting in use of every available memory channel).

[LuckyTran]

As far as I know, there is no limitation on the number of compute processes that Fluent can spawn even on the "local" configuration. Can you give a reference for the 32 node limitation? For example I can spawn any number of processes even on my quad core. Do you have a can't spawn more than 32 problem or can't compute on 32 problem? If you cannot stably compute, this reflects a problem in your setup such as mpi, etc.

Even as of Windows 10, the maximum number of physical processors that windows can support is 2 for most of the non-server Windows OS's. Are you running an OS that supports more than 2 physical processors?

Also, you generally cannot manually pick which CPU tasks spawn on. This decision is handled by the job scheduler (because of MMX/SSE instruction set, I forget which affects this property specifically).

[ettore]

Even as of Windows 10, the maximum number of physical processors that windows can support is 2 for most of the non-server Windows OS's. Are you running an OS that supports more than 2 physical processors?

I have already cleared this doubt in the first post:
''I should also mention that the OS is a Windows (that can see and use the 4P, 128 GB hardware). Other programs use / benefit of the complete hardware."

As far as : ''Also, you generally cannot manually pick which CPU tasks spawn on. This decision is handled by the job scheduler (because of MMX/SSE instruction set, I forget which affects this property specifically)""

I read here on an older Thread that under Linux this possibility exist. I was wondering if there is somebody that knows how to do it under Windows.

As previously said i found an publication/performance study from Ansys that says that after 30 processes on a local machine (aka computer) the performance gain drops and flattens at about 32 processes, that replicates exactly what i have experienced. In order to have more than 30 parallel processes you have to use at least 2 machines (PC's). After 3-4 PC's you have to use infiniband to gain performance.

Thank you for your opinion.