[Astan]

Hello everyone, I wanted to kindly ask you for information.
In these days I have to perform CFD simulations and to perform them in a more "efficient" way, I thought to perform the calculations in parallel creating a cluster with the PCs that I have available.
Unfortunately it is the first time that I interject myself with this necessity and therefore, being inexperienced in the computer field, I would kindly need, for information (just to understand if it is something I would be able to realize or if you need advanced computer skills) some straight ( without taking anything for granted, even on the most basic concepts) on how this calculation procedure "works", how to connect the PCs, in particular to understand a little the logic of this approach and what I need to do physically to make the connection .
I have a dualboot laptop with ubuntu 16.04LTS and windows10, a windows10 desktop pc and a windows10 laptop, and I have a modem with four ethernet inputs.
The simulations will have to be done with OpenFoam, so I guess the first step is to make sure that the remaining PCs are dualboot ubuntu, windows.
I thank in advance anyone who will read this post for the time dedicated to reading and also for any answers.

[flotus1]

Without knowing the specifications of the machines you have available, it is difficult to judge whether this project is worth it. But my initial guess from your vague description is "no".
You will have spend a substantial amount of time to get it working properly, and in the end with a heterogeneous cluster and gigabit ethernet connection the results might be rather poor.

[Astan]

Hello flotus1, first of all thanks for the reply.
Which specifications do you need in particular?

[flotus1]

CPU and memory, as precisely as possible.
So not "I5 and 8GB RAM" but instead "I5-4460 and 1x8GB DDR3-1600"

[Astan]

1) Hp envy dv6 - 7320el
cpu intel core i7 3630qm
memory 1 slot 4GB A-Data technology 1600MHz DDR3
1 slot 8 GB samsung 1600mhz DDR3

2) Hp desktop 110-500nl
cpu intel core i3 4160
memory 1 slot 4GB samsung 1600MHz DDR3

3) Hp notebook 15 ac618nl
cpu intel core i5 6200U
memory 1 slot 8GB A-Data Technology 1600MHz DDR3

[AliE]

Hello Astan,

What you are planning to do is surely interesting and I thougth to do the same thing "for fun". However, as flouts1 said, it is hightly probable that it does not worth in terms of performance. Modern clusters connect the nodes together using high speed infiniBand or something similar. This tecnology is much more than a modem with ethernet and beyond the hardware you have to deal with the software ( mpi and communications protocol). Do something similar is surely possible but about the performace well... you may spend 1 moth or more to assemble everithing and at the end your simulation will run faster on a 4cpu laptop. If you need compuational power, you can try to look at amazon cluster, cpu time costs few cents/h or you can buy a refurbished workstation if you do not want to invest to many money.

AliE

[flotus1]

Seeing the hardware now, I am tempted to confirm my initial guess of "no"
If you want results (i.e. faster solution times), it would be better to sanitize one of the machines and run your jobs there.
Either swap out the 4GB DIMM in your I7 laptop for another 8GB DIMM.
Or upgrade the desktop PC with a cheap I5 CPU and 2x8GB of RAM.
Setting up a cluster can be fun if you are into that kind of stuff, but it will be a hassle and results will probably be underwhelming with such inhomogeneous hardware.

[Astan]

Thanks for the reply to both AliE and flotus1.
I understand, so the efforts to generate a cluster could give not so good results.
I read that the PC with 12GB of memory is expandable up to a maximum of 16GB. Does an increase of 4 GB of memory result in major benefits in terms of computing hours from your point of view? would you advise me to increase it?
Moreover, even for information (maybe for future use), by chance, do you know very simple guides for creating clusters?

[AliE]

Hello Astan,

If you put more ram then your pc will be able to handle a bigger amount of data (e.g. bigger meshes). The increase of speed is related to other features of the memory as its frequency for example. Now memory is pretty cheap, you can buy 8gb for 50euro or less. If you decide for the upgrade pay attention on which kind of memory is already installed and buy 4gb with the same features. On you tube there are some guys which explain you how to build a server. Start from there

[Astan]

Thanks for the kind answers, I would have a question.
Do you think pc 1) could support a cfd calculation on a car using a 20 million cell grid? If so, how many hours of calculation might be necessary?

[AliE]

Hi astan,

I would say no. 20million is a fine resolution and with 12gb of ram it is possible to go out of memory. But regarding this point it depends a lot on the code you are using. Maybe the simplest way to find it out is try to run
. Btw, 4proc (5million per core) are surely not enough. I think you need 16 to 32 proc for this kind of simulation...
With a standard home pc you can run something arond 2million to be optimistic. Sorry

[flotus1]

Probably not. This will be too little RAM for most solvers and grid generators.
Solution time...somewhere between 10 and 10000 hours depending on what exactly you want to do

[Astan]

Thanks again for your kind answers and for your patience AliE, flotus1.
As you may have guessed, I am totally new in the field of simulations.
From your point of view, if I wanted to do a CFD simulation of a car similar to a formula 1, considering only wings, car body and wheels, (without considering the various aerodynamic details and complicated surfaces) as far as I could go with the pc 1) in terms of cell number?

[AliE]

For a steady simulation using SIMPLE family alghorithm, I would say within 2 and 5million, but I am much more for 2 than 5. Maybe flotus1 has a different experience and can be less conservative

[flotus1]

It really depends on the solver type you are using. If we take Ansys Fluent for example, its SIMPLE solver in single precision requires just under 1GB per million cells. In their training, Ansys allegedly recommends using the coupled solver for external aerodynamics. In double precision, this requires somewhere around 3.5GB per million cells. Since I never really used OpenFOAM, I don't know how the number are there. But you can find this out easily by running a smaller simulation yourself.
You can certainly simulate external aerodynamics of a race car with 1 million cells or less for training purposes if the geometry is simplified. Typical simulations for this kind of flow where you need actual results are 1-2 orders of magnitude larger.

[Astan]

Thanks to both for the answers, they are really precious information for me.
So using Fluent I could for example perform a single-precision simulation with 3 or 4 million cells, just to do a test.
Could you suggest me on which fluent tutorial can I find information on this topic?

[AliE]

Hello astan,

I do not know exactly since I do not use fluent very much, but since this is an incompressible flow problem, I think you can find something in the first pages of the ansys tutorial. Setting up fluent is pretty simple if you know what you are doing the most of the work is on the mesh.

Good luck!

[Astan]

hello AliE and flotus1, thank you for the availability and for the very important answers and advice. I wish you a Happy Easter