[Julian121]

I am running CFX simulation of a transient problem with 5M elements.
The simulation is too slow even though I am running it with parallel processing.
It takes an hour to complete about 30 time steps. Each time step has 5 or 6 iterations.
I have just found out that under host memory information out of 32 GB ram only 510 (1.56%) has been allocated at the start of run. Again, there is another memory allocation data which says out of 32 GB ram only 959 (2.94%) has been allocated. At last, there is another host memory information which shows Npart is 20 and 15GB (47.14%) has been allocated.
I have a workstation with 2X v3 xeons and am using 1 16 GB ram per cpu. The rest 6 slots are empty.
I wonder is there a problem that makes the simulation too slow and would I benefit from upgrading the memory?

[flotus1]

It's the memory population. You need 4 identical DIMMs per CPU.
So yes, upgrading memory is the most urgent hardware change you need.

[Julian121]

Originally Posted by flotus1

It's the memory population. You need 4 identical DIMMs per CPU.
So yes, upgrading memory is the most urgent hardware change you need.

Since 16 GB ram is expensive in my country, a seller recommend me to get 2*32 GB instead of my current 2*16 GB. He said that the performance would be equal to 8*8 GB. Is it true?
Does it really make huge difference when all slots are occupied?
Is there a big difference between 8*8GB and 8*16GB in my case?

[flotus1]

Your seller either does not know what exactly you are doing with your workstation, or he does not know what he is talking about.
For parallel CFD, memory bandwidth is what matters the most. Your CPUs have 4 memory channels each, and you need to populate each channel with at least one DIMM, and avoid using different size DIMMs. So you need 8 identical DIMMs, since you have two CPUs.
For performance in your particular case, it almost doesn't matter if you use 8x4GB, 8x8GB, 8x16GB, 8x32GB or even 8x64GB. There are minor differences, but since each of these configurations gives you enough memory to fit your simulation, these differences pale in comparison to bottlenecking the CPUs with only one memory channel.

Does it really make huge difference when all slots are occupied?

Yes, it does. That's why I recommended upgrading the memory in the first place.

Since 16 GB ram is expensive in my country, a seller recommend me to get 2*32 GB instead of my current 2*16 GB

I have trouble wrapping my head around this one. Are 32GB DIMMs cheaper than 16GB DIMMs in your part of the world???

[Julian121]

Originally Posted by flotus1

Yes, it does. That's why I recommended upgrading the memory in the first place.

Thank you very much for your advice.
I have populated all the slots with 8x16GB, but the simulation time did not change too much.
Before I use additional memory, it took about an hour to complete 12-15 time steps (each time step has 5-6 iterations). Now, it completes 25 time steps in an hour, but I was expecting a better performance. Is it normal for a transient problem with 5M elements to take too much time to solve?
Task manager shows that only 21% of memory is used. I tried to increase memory allocation factor but it did not change the percentage of the memory usage. Shouldn't the memory usage be more than this percent?

[flotus1]

Originally Posted by Julian121

I have populated all the slots with 8x16GB, but the simulation time did not change too much.
Before I use additional memory, it took about an hour to complete 12-15 time steps (each time step has 5-6 iterations). Now, it completes 25 time steps in an hour

Based on these numbers, you got a performance increase between 66% and 108%. Nothing to scoff at if you asked me.

Originally Posted by Julian121

Is it normal for a transient problem with 5M elements to take too much time to solve?

Maybe, but quite impossible to tell without knowing what kind of simulation this is exactly.
For a bog-standard single-phase simulation, solution times would be on the slow side.
For a multiphase fluid-structure coupling on the other hand...

Originally Posted by Julian121

Task manager shows that only 21% of memory is used. I tried to increase memory allocation factor but it did not change the percentage of the memory usage. Shouldn't the memory usage be more than this percent?

First of all: increasing memory allocation factor has no performance impact. Simply using more memory does not speed up most algorithms. For CFX specifically, MAF is more like a workaround, in case the standard value is not enough for some reason, and causes the simulation to crash.
Also, increased memory allocation does not show up as used memory in windows task manager. Glossing over pretty much all the details: task manager shows used memory (data that actually resides in RAM), not allocated memory. At least with default settings.
In order to make people with a background in CS completely flip their shit, an over-simplified example: You could make a program allocate 1TB of memory. If all the data of the simulation fits into 20GB of memory, task manager will show 20GB as "memory".

There could still be other factors holding back your simulation speed.
Hardware problems, software/driver issues, unnecessary disk-I/O...