This application forms a little stiffness matrix from a 
2D grid of springs.  The finite element formulation is
undoubtably horribly wrong, but this should be enough to
demonstrate how to solve matrix-based problems with real
boundary conditions using the FEM framework.

-------- Scaling Test on Cool -------
A 100x100 grid (10K elements, 30K DOF) can be solved 
serially in about 500 iterations, which takes some 15s.
This tiny problem doesn't scale up very well--4.7 seconds on
4 processors.  This is probably because the mesh setup
overwhelms the actual calculation.

A 200x200 grid (40K elements, 120K DOF) takes 1000 iterations
and hence 34 seconds (about 8x longer) on 4 processors.
The scaling is a bit better:
  +p4, +n4: 34s
  +p8, +n8: 18s
  +p8, +n8, +vp16: 18s 
  +p8, +n8, +vp32: 19s
  +p8, +n8, +vp64: 21s
  +p8, +n2: 47s (implies memory bandwidth is an issue)
  +p8, +n2, +vp16: 42s 
  +p8, +n2, +vp32: 42s
  +p16, +n8: 15s (counting bandwidth limit, this is OK)
  +p24, +n8: 14s
A single timestep is in the dozen-millisecond range, so on
this timeshared system we can't hope for much.

A 400x400 grid (160K elements, 0.5M DOF) takes 2000 iterations
and hence 143 seconds, which is right in line.
  +p2, +n2: 526s
  +p4, +n4: 296s
  +p8, +n8: 143s (perfect!)
  +p16, +n8: 110s