Hello Alexis,
I managed setup the environment that enables running a computation between two nodes but the results were disappointing with Ethernet, so I wouldn't recommend it. As the core count per node increases, it is getting pointless to spread computations across multiple nodes, unless the domain is really large.
For the record, I am using MS-MPI and HPC Pack as a job scheduler, but it is not necessary to have one. MS-MPI is the successor to MPICH2 for Windows and pretty mature by now, so I don't know why anyone would still use it.
If you really need MPICH2, there is little I can do.
Best Regards,
Costas