[jxs832]

Hi,

I am currently running a 3D transient simulation using rhoCentralFoam in parallel. My current computer has 4 cores, at 4.8GHz (overclocked), and 16 GB of ram (it only uses about 8GB during the run). I am running this in parallel and it takes about 100 hours or more to converge.

I recently got the money to build a little cluster, and I have seen some success in creating a cluster out of raspberry pis or similar boards. My question is what would be the better solution for my case. With the funds I have available I can make the following system:

- 34 nodes each with 2.1 GHz quadcore and 2 GB of RAM
- 76 nodes each with 1.6 GHZ quadcore and 1 GB of RAM

I am not sure how the computational time would scale with number of cores and clock times.

I should mention that I am using the scotch option in my decompose file, so theoretically each core is handling the same amount of nodes. Eventually, I want to increase my simulation to have a few million nodes.

[pbrady2013]

Hi,


In my experience there is really no right answer in this case as scale out results depend not only on your hardware - don't forget the network interconnects are just as important as the CPUs - but also the compiler optimisations and mesh/solver.


My personal preference is clock speed over count but I've no solid evidence to back that up, just intuition.


Ideally, you'd define the type of problems that you want to run and the test on some sample hardware before really scaling out. Then, to really squeeze performance, you'd want to custom compile your main compiler, followed by your MPI implementation and possibly look at your network stack and kernel.


Cheers,
-pete

[RobertHB]

Originally Posted by jxs832

I am not sure how the computational time would scale with number of cores and clock times.

Here is a scalability study of OpenFoam.