Is there an example of coprthr MPI spanning multiple hosts?

[capnrob97]

[capnrob97]

I changed it from displaying the hex digits to doing a graphical random walk on 150,000 hex digits of Pi.

Once I get the Epiphany cores working on this too (this was strictly done with the ARM processors), I will shoot for 1 million digit random walk.

[capnrob97]

Got this running on the Epiphany cores, and distributed to all 64 in my little cluster.

The performance is slower than I expected, got to tune the eCore code a bit, but glad I got finally all 64 cores working together on a problem.

[capnrob97]

100,000 digits with all 64 epiphany cores.

The pi digits stop tracking the real pi digits at digit #556, I assume due to having to convert the doubles to floats to run on the ecores.

Was faster though, this took about 1.25 hours, doing 125,000 just on the ARM processors took about 4 hours.

I am not going to pursue this one anymore, but now know how to pull all 64 cores together for a common task.

High level details:

I create 8 MPI tasks, one on each of the 8 arm processor cores in the cluster (2 per board)
Each MPI task creates an 8 core workgroup on the epiphany.
Each workgroup calculates its 8 digits of pi.
When each workgroup is finished, the MPI processes return the 8 digits as a string back to MPI task rank 0.

MPI task rank 0 gathers all the values and puts them in the proper order then draws the graphics for those 64 digits.

Rinse and repeat.

[mjvbhaskar1000]

[nickoppen]

Rob is not active on the parallella any more.

His nbody system is available here: but he did not post this system. You need a hdmi enabled parallella to see the output.

The best mpi examples are here: . These are the original versions of the mpi examples that come with the image. It seems though you have already found them given your other post.