On Feb 27, 2020, at 2:32 PM, Mark Davis via discuss <discuss@mpich.org> wrote:

A follow-up question: it seems that MPICH creates a duplicate MPI_Comm
for each window, as I mentioned above. Is that something that's
enforced by the standard or just an MPICH implementation detail?

That is implementation detail. However operations on different window is meant to be independent — ie can happen concurrently. In the case memory exposed in different windows overlap, the MPI standard state that the results are undefined.

— 

Hui Zhou