
3)  Make output of all test cases deterministic, so we don't get spurious
    diffs.

4)  Translate the other 15 test cases in

      /project/pss_shko/mpi_TestSuites/IBM_mpich_test/dtyp

    to Java.

7)  Translate missing cases in `IBM_mpich_test/pt2pt'.

8)  Make sure `pt2pt' tests for multiple completions are sane.  Currently
    some are run with request vectors with only one request.

9)  In the "MPI for Java" document I wrote:

      "For JNI-based ``wrapper'' implementations based on the C binding
      of MPI, one should ensure that all MPI functions, including the
      `MPI_..._free' object destructors, are called *before*
      `MPI_Finalize' is called.  Because invocation of Java destructors
      is under the control of the garbage collector there is no guarantee
      that all Java-level `finalize' methods will be called before the
      explicit call to the Java-level `MPI.finish'.  A possible solution
      is to maintain a global count of the number of outstanding
      `MPI_..._free' calls.  This is initialized to 1 by `MPI.init' and
      incremented by relevant constructors.  All relevant Java `finalize'
      methods and `MPI.finish' should decrement this counter, and be
      prepared to call the physical `MPI_Finalize' function when and only
      when the count falls to zero."

    We don't actually do it.

10) Add proper exception handling!

