*******************************************
            *** Parallel Programming with MatlabMPI ***
            ***         Dr. Jeremy Kepner           ***
            ***      MIT Lincoln Laboratory         ***
            ***        kepner@ll.mit.edu            ***
            *******************************************

TO DO LIST

-Demonstrate an interactive server.
  examples/server
    server.m
      rank~=0 spins on eval(MPI_Recv())
      increments tag each time a message is received
    multEval('command',ranks)
      sends command to ranks, default is everyone.
    var = getVar('VarName',ranks)
      multEval('MPI_Send(comm,0,tag,'VarName')');
    putVar('VarName',ranks)
      sends two messages
        MPI_Send(comm,dest,tag+1,'VarName')
      multEval('VarName = MPI_Recv(comm,source,tag+1)')

-Explore performance benefits of using a Ram disk

-Eliminate need for user managed tags by keeping track
 inside comm.
  -Comm keeps a list of recv and send counts for each
   processor and each tag.  Increments after save/load. 
  -Drawback: requires that comm holds state info,
   lack of pointers or rhs modification means a fundamental
   change to syntax:
      [comm] = MPI_Send(comm, ...);
      [comm, vars] = MPI_Recv(comm, ...);
   or use of globals, which also changes syntax.
   Implication: every line of MatlabMPI code will need
   modification.  Also limits to tags to fixed number of
   integers.
   Recommendation: do not implement.

-Allow comm scoping?  Only makes sense if doing tag
 management (see above).  Limiting tags to integers
 prevents user scoping and is incompatible with pMatlab.
   Recommendation: do not implement.