Data distribution:

          A                        B                         C

   +-----------------+        +---+---+---+---+         +-----------------+
   |      P0         |        |   |   |   |   |         |      P0         |
   +-----------------+        |   |   |   |   |         +-----------------+
   |      P1         |        | P0| P1| P2| P3|         |      P1         |
   +-----------------+        |   |   |   |   |         +-----------------+
   |      P2         |        |   |   |   |   |         |      P2         |
   +-----------------+        |   |   |   |   |         +-----------------+
   |      P3         |        |   |   |   |   |         |      P3         |
   +-----------------+        +---------------+         +-----------------+

--------------------------------------------------------------------------

Potential code structure:

for (i=0; i<nprocs; i++) {
  Node i sends its B array to all other nodes
  All other nodes receive the block

  Call mm_sse(localN, M, LocalK, a, receivedB, workC)

  Copy workC to C
}

Some notes: 

1. You need to send the whole block in one shot. Sending smaller units take
   more time.
2. You need to allocate working space (e.g. receivedB and workC). ReceivedB may
   be bigger than the B chuck that a node has.
3. Ctrl-C to may not kill an MPI program gracefully, make sure you take care of
   the leftover processes.
4. mpdboot once and use forever.
5. There is a limit on the number of processes that you can have on each 
   machine. Running MPI programs can easily overrun that limit.