Using python for multi-GPU simulations with GPU direct

[syoffe]

Hello,

I was wondering if anyone has had any success using GPU direct with python? We managed to get a version running using mpirun from openmpi-gcc5/3.0.0 but this was significantly slower than using mpirun from our Anaconda installation without GPU direct. Is there a good way to set up python for multi-GPU simulations using GPU direct?

Many thanks for any help/advice,
Sam