Non-blocking object send/recv?
    Currently only offered for arrays
    Need to pre-allocate recv buffer may be a problem
Many more tests
    buffer allocate / deallocate
    send modes
    etc
Topology
    Dims, shift etc.
MPI_*v collective methods
    eg: MPI_Gatherv rather than using gather(..., concat=True)
MINLOC, MAXLOC
    New numpy array types may make this easier, otherwise use 
    temp arrays or limit datatypes
Datatype restrictions for reduce operators
    MPI spec says some datatypes not appropriate
Vampire build in setup.py
    For profiling on sc
Non-contiguous arrays
    If MPI new-datatype overhead is low enough, should be
    able to send/receive from ANY Numeric array as is.    
Intercommunicators
