I'm writing my thesis and i have built a RPI cluster, containing 10 nodes which consists of RPI model 3b. I've them connected to two gigabit switches. I don't know the CAT of the cables. They are not connected to the Internet, they just live in their private network.

Further more, i've calculated the theoretical performance by the formula:

number of cores * average frequency * 16 FLOPs/cycle = x GFLOPS (Got the formula from https://www.slothparadise.com/how-to-run-hpl-linpack-across-nodes/ )

after i applied it, it turns out i should have 76,8 GFLOPS in theory assuming that the source is relaiable and correct. When i benchmark it using HPL 2.1 i've only reached little about 7 GFLOPS at best when trying out various variations.

Now to my question: i've read up on the RPI model, and it says it only have 10/100 Ethernet. Is that the source of my problem? Seeing i get out GFLOPS from a node, but the network transportation is much less then Gigabit speed which would mean 76,8 GFLOPS * 0,1 Gigabit/second = 7,68 GFLOPS (in theory). Or am i way off the tracks in my thinking?

I really appriciate any help, so i can know if i have to keep working on the configuration for the benchmarking or if i can move on. Also, i'm sorry if i have posted in the wrong place.

Stay safe out there!

  • $\begingroup$ Are you confident a formula for Intel Haswell/Broadwell/Skylake applies to Raspberries? Or nominal network bit rate relates to floating point throughput? $\endgroup$ – greybeard May 2 at 18:51
  • $\begingroup$ i was not thinking, see my answer down below. Thx though :) $\endgroup$ – mozzie May 2 at 20:02

I think i have solved it now. The problem was that in the example in the source they had:

8*3,50*16 and i was assuming (which is a very bad thing to do) that i should change everything except the 16, that it was a constant of sorts. Turns out, my RPI should have somewhere between 0 and 2 in the last multiple, alas it should be

4*1,2*x where x is somewhere between 0 and 2 (as my estimation) thus i have only to find out the actual number for X, and for some reason it seems hard to find it, one would have thought that the instructions per second or instructions per cycle would be documented for RPI.

bottom line, there was nothing wrong except for my thinking. sorry for this post, i should had read up a little bit more before i posted. I blame it on stress.

EDIT: i forgot to give the source that helped me solve the problem, here it is:


and thx for the help!

| cite | improve this answer | |
  • $\begingroup$ Good of you to identify sources. $\endgroup$ – greybeard May 2 at 20:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.