open TELEMAC-MASCARET The mathematically superior suite of solvers

Hi all,

I could have the chance to build a cluster of 4 RaspberryPi3 (1.2 GHz Quad-Core ARM CPU). This would result in having 16 computation cores. As shown in the Benchmark section, it is generally possible to use the Pi for Telemac computations and I assume that it would also possible to get it to run as a cluster.

I am currently computing on a workstation with a Xeon E3-1241 (2 physical cores @ 3.5 GHz).

Now the question - would the larger amount of cores in such a cluster significantly speed up computations - even though the single cores have only 1/3 of the processor frequency of the Xeon? After all it's 2+2 cores Xeon vs. 16 cores Pi.

I would be happy to get some insight on this topic / find out if it could make sense before actually spending the time getting the Pi-Cluster to work...

Hi,

Not exactly the answer you are looking for but surely this topic can be helpful.

Regards,

José Díaz.

Hi,

interesting idea. I'm not an expert but i guess that the a RaspberryPi cluster is more a playground.

You can look at results of the RaspberryPi benchmark section and compare the results with your machine running the same benchmark.

The network interface of the RaspberryPi will be limited compared to a real HPC cluster. The Intel CPUs can use vectorization,.... (depending on the code) so you should not only look on GHz. Further you must check how good the compiler will optimize the code for the RaspberryPi. The memory bandwidth will be also different.


Best regards

Leo

Thanks José and Leo! In the meantime I did some more research. A rather large slowdown seems to be the memory bandwidth of the ARM-Architecture which is used for the CPU of the RaspberryPi. This is apparently a major restriction/slowdown for CFD computations on the Pi.
What I took away from the threads in other Forums is that from the standpoint of computation time it does not make much sense to build that cluster - if you have some more or less powerful x86/x64 CPU available. So it would be only "useful" as a learning experience on clustering.

Might be a nice project for learning, but for production I'm going to stick to my Xeon for now.

(However, Uwe's post in the thread linked by José is suggesting that it might make sense to use mroe CPUs for larger problems - that's mostly what I have on my table. So maybe I can get some boost with it? I'll make sure to post some benchmark here, if I build this cluster and/or get it to run).

We run Telemac on a cluster, most people use between 1000-10000 FE-Nodes per MPI rank. So most TELEMAC2D models run with something between 80-240 MPI ranks.

I don't believe that you can get any interesting boost with 4 Raspberry Pis. But nevertheless it is interesting to know the difference between INTEL and ARM systems. I think there are better ARM boards for that purpose.