Computer power

[Luke]

Hi. I am looking for some help in choosing a new computer specifically for running FDS on at work.

We currently use two or three old 32 bit dual core Intel Xenon at 4 GB machines (running Windows XP). We use parallel processing across them. We are interested in getting a new 64 bit Six Core model with a lot of RAM. We need to shorten the run times on our models.

Typically, we would run models with under 10 million cells (more like 5 million normally). Would the new machine significantly shorten the simulation time? As in, by at least half or three times compared to the old machine? Or would it take the same amount of time to run the simulation unless I split the model into more meshes?

We also need to decide how much RAM to get. Is RAM related only to the amount of cells you have? For example, if I had a 10 million cell model and a computer with 10GB RAM, would it run faster on the same computer but with 128GB RAM?

This new computer will only run FDS models, so if anyone has any advice as to what the best option is, it would be greatly appreciated.

Thanks for your help.

[F-Sim.de]

> Would the new machine significantly shorten the simulation time? [...] Or would it take the same amount of time to run the simulation unless I split the model into more meshes?
Using a single MESH the new macine will be faster significantly (using OpenMP), but I would not expect a factor of 3. Using MPI, compared to your 32bit Xeons (> 8 years!), you can be 3times as fast.

 

 > Is RAM related only to the amount of cells you have?

Yes. It depends a little on what cells you have (wall cells f.e.) but roughly 500 MB is needed for 500.000 cells.

 

> would it run faster on the same computer but with 128GB RAM?

All the cells need to be stored in RAM, no file swapping should be done. Above this more RAM will not speed up your calculation significantly.

 

Buy a new Xeon or Core i7 with as much cores and RAM as possible. Use a x64 OS.

 

Boris

[Peter_J]

Hi Luke,

I have ran several large-cell cases (30-50mln cells) last months, and so far my conclusion are as follows:

1. When you are running single mesh case, you can assume: 1mln cells / 1,0 GB RAM
2. When you are running large scenarios involving use of multiple parallel meshes: 1mln cells / 2,15 GB RAM (calculated from my largest case).
3. It is crucial to remember about some memory reservation for Your OS. This value can differ between 0,5-2GB depending on Your installed OS

If You are building model where amount of grid cells is close to Your RAM capacity - You cant be sure, if solver will use your hard drive to dump some temporary data from calculation process (page file / virtual memory)  - or not. If it happens, You will notice significant performance drop of Your calculation. So, If You are performing calculation with number of cells close to Your maximum RAM capacity, and then You expand your memory - you can observe some improvement in calculation speed - all this can happen because solver will stop using virtual memory. But further expansion of RAM wont bring any difference. So, discussion about impact of amount of RAM in terms of callculation speed - can be reduced to removing bottleneck problem. If Your amount of memory is adequate the main role in calculation speed plays speed of Your cores, and of course: number of cells per core ratio.

It is good practice to evaluate maximum amount of cells in accordance to rules of thumb listed above, before starting any calculations.

In my opinion You should have two times more GB of RAM, then the cores if You are going to run large sceneries.

I have single calculation station, equipped with 4-processor mainboard with 16-core OPTERON CPUs. It gives 64 cores on single work station. I have 64GB of RAM, but according to my benchmarks this value must me multiplied by 2 in order to use full potential of this machine. In next Year I will run several benchmarks to study more deeply problem of calculation performance against memory used, optimal cells amount per single process and other problems - and I will gladly share results with FDS community.

[Erichovi]

Very interesting observations indeed! Looking forward to hearing about the outcome of the simulations you will run within the next year.

[Monika]

Hello peter

I have the same problem with choice computer parameters to FDS. Maybe you know:

1. Should I have minimum 12x2= 24 GB RAM for processor Intel i7 (6 cores 12 threads)? Does the FDS use a lot of threads in cores?

2. Does the FDS use to calculate GPU - CUDA? Does the GPU influence to speed of calculate?

3. I have purpose to use computer program PYROSIM?

 

Thank you in advance

Monika

[Evgeny Bessmertny]

Hi Luke!

Which kind OS  best match for reduce  calculation time of FDS? 

понедельник, 22 октября 2012 г., 17:03:32 UTC+4 пользователь Luke написал:

[Peter_J]

Hi Monika,

1. I havent done any benchmarks with use of HyperThreading technology, but assuming the same node/RAM factor computed from my last simulations I would strongly recommend to include 24-32GB of RAM into your workstation.
2. From my knowledge there is no available CFD code that uses GPUs for calculations, so FDS do not use it neither. As far as I know, calculations with multiple GPU units are implemented in some of ANSYS software modules dedicated for mechanics/structural simulations, but not CFD. At present state of CFD codes development, only CPU speed, number of nodes and RAM matters.If I am mistaken I hope somebody will correct my statements.
3. Using Pyrosim or not - do not change anything in that matter. Pyrosim is a graphic interface. FDS is used as a physics and numerical solver, and MPICH / or OpenMP is used for parallel runs.

Best regards
P.

---------------