While the discussions of in the last several sections have hopefully illuminated and clarified the landscape of MPI implementations, the overall picture had been a bit unbalanced, because of the natural tendency to focus on what doesn't work, rather than what does. The primary message you should carry away is still that there are many MPI implementations out there, that they are well-supported by vendors, and that despite the occasional bug or missing feature, they work quite well. For instance, while I have mentioned that several implementations are not MPI 1.1 compliant with respect to the definition of certain Fortran constants, the overwhelming majority of the MPI standard has been correctly implemented by all implementations.
The area in which there remains the most room for improvement is integration of MPI applications into a robust environment in which parallel applications are first-class citizens, understood directly and handled appropriately by the operating system, debuggers, analysis tools, and user interface.
Also high on the MPI wish list is a freely available, robust and highly usable MPI for networks of workstations (NOWs). While MPICH (through P4) and LAM fill some of this need, they have some usability problems as discussed above. What is needed is something with the process and virtual machine management of PVM, but with the MPI message passing API. This type of software is needed both at the user level (so that users can put together personal NOWs) and at the system level (so that administrators can build highly integrated MPP-like systems).