How many GMM can MapReduce do? What about lattice quantum chronodynamics performance? MPI and Lustre exist for a reason: map reduce isn't great for all problems.
MR never claimed to be great for all problems. It main selling points was big-data and easy-of-use.
Sure, MPI might blow MR out of the water in term of number-crunching, but it is also way harder to use.
Without Google using, validating and releasing those design, we might be stuck with MPI and NFS for a lot longer.