yade-mpi team mailing list archive

[Bruno Chareyre <bruno.chareyre@xxxxxxxxxxxxxxx>]

(re-sent, correct adress to yade-mpi is yade-mpi@xxxxxxxxxxxxxxxxxxx)
Hi there,
An effective N-threads implementation of distributed memory parallelization
is achieved after minor changes in the source code and some refactoring of
the previous 2-threads script.
You can find the concrete example testMPIxN.py with comments in [1].

The 10x slow down I mentioned previously was due mainly to triggering
contact detection based on false positive. The monolithic run was executing
collider only once (correct), the parallel version was triggering detection
all the time (superfluous).
At the moment the mpi version is on par with monolithic version for 90k
bodies (and 3 subdomains for the parallel version). This is more than
encouraging, but note the geometry of the example simulation is very
pleasently parallel.

I cease development at this point and leave it to future works. I am more
than open to discussions though, since there is very little documentation
and questions/answers might help to explain what happens therein.

Cheers

Bruno

p.s. this is the very-very last email to yade-dev, thank you to those
registered to yade-mpi  :)


[1] https://github.com/bchareyre/yade-mpi/commit/
fa48e931cf44c12ed78f288f931f0981955635d7