search

cwpearson / stencil

A prototype MPI/CUDA stencil communication library
Boost Software License 1.0
10 stars 3 forks source link

Wrongly assume that all ranks share the same GPU topology #7

Open cwpearson opened 4 years ago

cwpearson commented 4 years ago

https://github.com/cwpearson/stencil/blob/e821f0c6ac007b813d7a72c558c6d6a86179094a/include/stencil/stencil.hpp#L371-L372

If two ranks are on a multi-GPU machine, one rank will have GPU 0 and the other rank will have GPU 1. This code does not work for this scenario.