Each benchmark acts on a certain number of points for a certain number
of iterations.  Let me call the work that is necessary for one point
and one iteration an "event", borrowing this terminology from
relativity.  The number of events for each benchmarks is then:



Benchmark			iters	points		RHS	total events
================================================================================
Bench_BSSN_PUGH_80l		 80	  80^3		3	122,880,000
Bench_BSSN_PUGH_100l		 40	 100^3		3	120,000,000
Bench_BSSN_Carpet_1lev_60l	160	  60^3		3	103,680,000
Bench_BSSN_Carpet_1lev_75l	 80	  75^3		3	101,250,000
Bench_BSSN_Carpet_3lev_32l	 84+9	  32^3		3	  9,142,272
Bench_BSSN_Carpet_3lev_42l	 42+9	  42^3		3	 11,335,464
Bench_Whisky_Carpet_36l		224+9	  36^3		3	 32,612,544
Bench_Whisky_Carpet_48l		112+9	  48^3		3	 40,144,896

Bench_BSSN_Carpet_3lev_32l	 84+9	( 32+8)^3	3	 17,856,000
Bench_BSSN_Carpet_3lev_42l	 42+9	( 42+8)^3	3	 19,125,000
Bench_Whisky_Carpet_36l		224+9	( 36+8)^3	3	 59,543,616
Bench_Whisky_Carpet_48l		112+9	( 48+8)^3	3	 63,748,608

The column "RHS" counts the number of right hand side evaluations per
time integrator iteration.  Also, note that the mesh refinement
benchmarks iterate over the fine levels more often than over the
coarser levels; we count each level as one iteration.

The "+9" in the iteration column is a number of additional iterations
to create second order accurate initial data.  The "+8" in the points
column is a number of additional grid points which are placed on
refinement boundaries (but not on the coarsest grid [TODO: Correct the
total events column for that]).

The number of points was chosen so that the benchmarks occupy
approximately 400 MB and 800 MB per CPU, respectively.  The number of
iterations was chosen so that the benchmarks finish in about 15
minutes (on a fast CPU in 2005).

The BSSN benchmarks all solve the same set of equations, namely the
Einstein equations in the BSSN formulation.  The Whisky benchmarks
solve in addition the relativistic hydrodynamics equations; the Whisky
benchmarks also use 3 refinement levels.



The table above counts only operations towards solving the physical
problem.  In addition, there are operations to manage the
inter-processor boundaries, and to manage the mesh refinement
boundaries.  Especially the latter can be very expensive.
