Raspberry Pi (or similar) and FDS

[Matthew ..]

Hi All!

So I dont really know anything about super computers, cluster computing, cloud computing, distributed systems, linux, MPI, or any other technical details. However, I did read an article about many raspberry pi's being linked into a lowcost "super computer" which could use MPICH2 to calculate pi (I think there was a pun/joke in there somewhere)

So I was just wondering if this would have any benefits for FDS? From what I can gather using MPI then one mesh will be calculated on one device, so your speed would be limited by the processor of the device and the size of the mesh would be limited by the ram. Is that correct? Each device has only 256 MB ram, so that would severely limit the size of the largest mesh? Larger models may need many meshes, which may lead to numerical instabilities etc. The second problem is that the CPU is only 700 mhz, from what I understand. So that would limit the speed of calculation.

Would it be possible to use openmp using a system like this?

So, I am really just wondering if this type of system might be used in practical FDS applications? I guess a better solution may have more ram and a faster processor, but less of the extra stuff like usb ports and hdmi out etc.

Regards,

Matt

[shostikk]

What makes a super computer super? Usually it is the speed and bandwidth of communication between the (many) individual processors. For some applications, such as calculating PI using some distributed series, this is not crusial at all. For some, such as CFD using domain decomposition-based parallelization, it is very important. In the case of FDS, connecting many low-cost processors with cables would not be efficient.

Simo

[Franck Didieux @ LNE]

Dear Matt,

The problem is slightly different with FDS than for cryptography. In the last case, the problem to be solved can be divided in as many subsets as you wish, each of them being independant from all others. Then, the calculation may be performed massively in parallel : each processor or computer perform it's own calculation, blind about what others do and return the result for its subset once it has finished.
In FDS, the problem is a whole, that cannot be divided so easily : each subset has to transmit information to the others, and vice versa. Consequently, if you divide in to many subsets, the time you save in calculation is lost in transfering information. From experience, you can divide the problem up to a point that is dependant of the bandwidth for the communication between the processors (and of course the calculation speed of the processors).