Thanks for the word of caution! I've been using the remotecall functions rather than MPI (bunged my code into ClusterUtils.jl). Don't know how I'd use GPUDirec

Thanks for the word of caution! I've been using the remotecall functions rather than MPI (bunged my code into ClusterUtils.jl). Don't know how I'd use GPUDirect with this strategy. The packages based on CUDArt.jl all seem to work with remotes. So long as you don't fetch a pointer to memory on a remote device.