[cattewell]

Hi all,

I currently use Mentor Graphics FloEFD and I would like to upgrade my PC to speed the calculation time up. The current specification is:

Dell Precision T3500
Processor: Intel(R) Xeon(R) CPU W3550 @ 3.07 GHz
RAM: 12GB
System type: 64-bit Operating system

Please can you suggest what the best thing to upgrade is. The PC can be upgraded to a max of 24Gb of RAM but I am not sure if it would be more benificial to upgrade the processor.

Many thanks

Cameron

[akidess]

Depends on what you want to compute. FloEFD eats RAM like candy, but a four-core processor is also outdated. If you are currently fine with the fidelity and complexity of your models, update the processor. Otherwise update the RAM (won't really help with speed).

[cattewell]

Hi Anton,

Thank you for your quick response. I work for a company that designs and manufactures Automotive lighting. Therefore I am working with fairly complex assemblies which also sometimes requires me to include sections of the vehicle.

Hence why I am struggling with my current machine.
Firstly why will the RAM upgrade not speed the simulations up?
Secondly can you please recommend a couple of processors you would recommend?

Many thanks

Cameron

[CeesH]

why would it, unless you see your RAM is actually full (which you can check in task manager)? Or are there clear indications that memory bandwidth is lacking (and that this is due to the memory bank itself, and not the number of memory channels on the CPU)?

Before you start upgrading, it's good to determine what your bottleneck currently is - are the processors typically 100% loaded, or is the RAM full and are processors idling most of the time?

If you are looking at new processors, I have good experience with the AMD Ryzen options - they offer similar speed as typical XEON processors but at much lower cost. You can find some good, in depth suggestions for hardware in the hardware section of the forum, too.

[Boris_M]

Hi everyone,

There are good points from Cees and Anton.
RAM does not really influence the speed, except if you run out of memory and it starts using virtual memory on the hard drive as Cees mentioned. This can become very slow and you would see that in the CPU performance in the task manager when processors are not working full power.
Of course, there are RAM modules that might have a faster communication speed, but the processor will certainly speed things up more.

You should focus on the CPU as even every new generation can bring 20% speed increase plus of course higher clock speeds help as well. Of course the price is usually a limit as well but if you want to upgrade your computer with the existing hardware, then you face another problem of finding a faster processor that still fits to the motherboard.
However, I do not recommend AMD as the FloEFD development team is working closely with Intel to optimize the solver for Intel processors.

So in short, it is best to buy a new computer rather than update an old which has limitations given by the existing hardware (or you would need to replace the motherboard, CPU and likly also RAM for that old computer). Always pick as much RAM as you will likely need, I recommend at least 24 if not 32 GB to be on the save side and always pick the fastest Intel processor your are able to afford.
I usually check the ranking on www.cpubenchmark.net.
Here you find the list of all processors and can see where yours ranks compared to the latest models.

Be aware that some can be quite expensive but only have 8 faster cores compared to cheaper ones with 12 or more slightly slower cores. It depends on how many solver licenses you have available and how many projects you run in parallel. If you run 2 projects in parallel on 6 cores each, they will both be done in for example 6 hours and if you run them on 8 cores in series, they might be done in 4 hours each, but adding them up you will need 8 hours for the calculation.
So it depends on what your setting is that they can actually help you speed things up compared paying a lot of money.
If you have a standard supplier such as Dell or HP, check what CPUs their configurations offer and then compare them by speed in the benchmark list. The benchmark does not relate directly to FloEFD, but it gives an indication of the overall performance of the processors. A slight difference such as 25100 vs 25400 points in the benchmark don't make the much higher price worth it, but if you have a difference between 21k and 25k and barely any price difference, then it makes sense to consider the faster one, that is if the number of cores can be leveraged by you efficiently as mentioned above.

When I search for your processor, I find it at 5706 points of the benchmark which is Number 613 in the ranking compared to a new model such as the Dell Precision Desktop 5820 Tower with an Intel Xeon W2145 with 32 GB RAM for ~3500 $ where the processor has 19819 points and Rank 34. So the benchmark points are not quite 4 times higher.
Just as an example.

I hope this helps,
Boris

[Daniel C]

Hi Boris,
you say FloEFD it optimized for Intel processors. What about the new Epyc CPUs from AMD with 8 channel interface. Are there any experiences regarding these processors?


Ciao


Daniel

[Boris_M]

Hi Daniel,

Yes, FloEFD is optimized for Intel as their programming libraries are used. That doesn't mean it doesn't work with AMD, but it simply is not optimized to it's best as simply most business workstations run on Intel.

Boris

[Usemia]

Hi all,

What is the maximum number of cores that FloEFD will benefit from?

Will FloEFD benefit from a multi-CPU (muti-socket) system? Multiple cpu's can help with the memory bandwidth. How about a quad cpu or an 8 cpu server system?

[Boris_M]

It doesn't scale linear. So it depends on how many solver licenses you have and how many projects you could typically run in parallel to tell if additional cores give a good ROI.
For example, if you have 40 cores, FloEFD can use them, but after maybe 20-24 the increase in speed is only little that the additional CPU with 20 cores would not be worth it. But if you have two solvers, you can run two projects at the same time with each 20 cores and therefore double the speed to get the projects calculated compared to maybe 1.3 times faster from 20 to 40 cores and calculating them in series.

In numbers, if 40 cores are only 1.3 times faster (I don't know the actual number), then 2 projects that take each 4 hours on 20 cores will take about 3.08 hours on 40 cores and in series the two projects take 6.16 hours compared to 4 hours with two solvers in parallel.
As you can see, it might be worth to have several solver licenses rather than 100's of cores.

Yes, FloEFD can use all cores and CPUs in your single computer/node. As long as those CPUs are treated with the shared memory approach.
FloEFD can also run on clusters which are recognized as individual computers/nodes where a solver license can run a project per node, but it cannot distribute a single the calculation over all nodes, this would be the distributed memory approach.
https://slideplayer.com/slide/940481...ted+Memory.jpg

So it would depend on how fast your bus is that the CPUs can communicate with the RAM without any delay.

Regards,
Boris

[Usemia]

Hi Boris,

Thank you for the explanation.
It's seems that access to the memory is a key player here.
With modern Xeon cpu's, the memory controller is on the chip. Each CPU has DIRECT access to it's memory modules that it controls directly. It has access to memory controlled by other CPU's (within the same system/node) INDIRECTLY through the QPI link.

My question is, for a single calculation, would FloEFD perform faster on 1 physical CPU with 30 cores, or on 2 physical cpu's with 15 cores each? Both setups have a total of 30 cores but the way in which the cores access the memory is different. (Lets assume for this example that both CPUs are of the same generation or same IPC-instructions per clock)

[Boris_M]

As you said, it will improve the communication bandwidth with the memory, so it depends on how much communication is happening and how much that is relative to the CPU time to solve any simulation.
For example if communication with the memory is assumed to happen at the end of each iteration and it only takes 10ms but the iteration time is 1 minute, then it will barely change any overall CPU time with 250 iterations it will only be 2.5 seconds in total and if it will be 2 or 2.5 because of the separate memory controllers, it is not speeding up the simulation really. And since the memory usage depends on model dimensions it is hard to tell how much influence it will have. Now those numbers where just an example and I have no idea how much it is and the actual values don't matter as only the relative speed increase will matter. So yes, a faster bandwidth of the QPI and separate memory controller will certainly influence it, but I cannot tell how much that is as it would just like the CPU performance itself also depend on the physics that is used and other factors.

The problem is, that there are so many factors influencing the simulation speed and so many different processors and new processor generations coming out every few months, that it is impossible to keep up with testing them simply for performance benchmarks.
I think you are getting lost in a detail that will at the end barely matter, it is best to consider the overall picture.
It is best to use your budget limit and work from there to the components you need and use the rest of the budget on the CPU and the required components such as the fitting motherboard and RAM module types.
Once you are in that end phase of picking the processor, the choice is usually limited to a few high-end ones and the question is then if additional $4k are worth a 3% increase in performance if at the end it turns out you are mostly using heat transfer simulations and the high mach number solver scales better and is raising the average scalling numbers of the benchmarks that were done. Some physics simply scale better than others and the maximum scaling might also vary on the model tested and the average is on all models that were tested. So another uncertainty.
This at the end leads to an overall uncertainty of x% speed increase and probably equals any benchmark of the processors which can be found on www.cpubenchmark.net which have a high uncertainty for such new systems anyway as only little benchmark data exists.
So don't try to squeeze the quantum bits out of it, as it won't be worth the effort and you won't notice it at the end either.

What system are you currently using and how long do your simulations typically currently take?

Regards,
Boris

[flotus1]

Originally Posted by Usemia

Hi Boris,

Thank you for the explanation.
It's seems that access to the memory is a key player here.
With modern Xeon cpu's, the memory controller is on the chip. Each CPU has DIRECT access to it's memory modules that it controls directly. It has access to memory controlled by other CPU's (within the same system/node) INDIRECTLY through the QPI link.

My question is, for a single calculation, would FloEFD perform faster on 1 physical CPU with 30 cores, or on 2 physical cpu's with 15 cores each? Both setups have a total of 30 cores but the way in which the cores access the memory is different. (Lets assume for this example that both CPUs are of the same generation or same IPC-instructions per clock)

This boils down to one question: did the developer account for NUMA architectures.
Apparently they do since v10: https://blogs.mentor.com/robinbornof...el-cfd-solver/

Memory management is modified to take advantage of the NUMA-architecture (Non-Uniform Memory Access) of modern processors