Things to be careful about in the Jacobi MPI implementation
---------

1) logical to physical mapping

   Declare the meshes X and Y:
     X[N+2][N+2] and Y[N+2][N+2]
   how to partition?
     Logical array may be distributed in the first dimension or 
     the second dimension.
     X[N/P+2][N+2] 

     Each node would still want to have a size of dimension+2 to store
     the boundaries 

   How is the mapping done?
     X[i][j] --> X[my_id*pernode_size+i][j] (something like that)
     You need to work-out the index mapping, especially when the 
     matrix size is not divisible by P.

     Which code is affected? 
       Initialization
       Loop bounds (loop body is similar if boundary is maintained)
       Output

2) Managing communication:

   in the main loop:

     each node must send boundary to its neighbor and 
     receive data from its neighbor. 

     Be aware of the potential deadlock situation. When can this happen?
     See example

     Communicate Maxdiff: need to follow a reduce broadcast pattern
     
  in the output:

     Cannot let all processes write to the same file
     Common approach:
       Send data to process 0 and let process 0 write to the file only
       All processes can send at the same time, 
       but process 0 will receive in order.