We need to setup a coupled case. As a minimum we need four pieces of information to do this:
1) The CFD domain size 2) The MD domain size 3) The overlap size 4) Processor topology in both domains 5) Grid or cell information
Only 1) to 4) are required to create the map between processors in both domains. The CFD, MD and overlap sizes is sufficient to determine processes which will communicate. However, in order to send cell or grid based information, the physical cell locations are also needed to identify which part of a processor is send to another processor. There is only a single grid here -- specified from the CFD CPL_setup routine.
In CPL library, the current setup assumes a uniform grid. This is because this is the simplest approach to match regions in space between the CFD code and the MD regions used to bin data. Even with a non-uniform grid this approach makes sense as the overlap region could still be uniform or interpolated to exchange data.
The philosophy was that CFD processors would be much bigger than the MD so would never be a problem (most of the computer/processes are allocated to MD so more than one proc in y for the CFD was unlikely, let alone in the overlap). Similarly, the coupled cases assume there are multiple MD processes per CFD, and doesn't allow multiple CFD per MD.
If the constraint region is on one processor and the bottom of the overlap on another. Or alternatively the constraint region is split over two processor, it becomes more complex to send the data and manage the applied constraint.
If this is something that presents a problem then we can look into fixing (shouldn't be too difficult). It is mainly just lack of development for more complex cases (in the coupler mapping) which is designed to be extensible.