*** Timing Estimates for 3D Turbulence Runs ***
Mark Petersen, Oct 24 2003

I have run timing tests on a single processor Itanium 2 for 128^3 and 
256^3 grid points.  My code uses 7 large arrays, so memory requirements 
can be estimated and compared to the code.  Using the time per step from
these runs, I can project the time required for larger runs.

Grid points	Memory, MB	time/time step
Per side	SIZE		sec
  32		   6	
  64		  16		 0.8
 128		  96		 6.1
 256		 734		56.2
 512		5006		died
1024		Failed	

For my rundown experiments with the time scaled so that initial energy is
1, I can go to t=10 for the initial run-down or t=100 for the full run to
two consolidated vortex columns.  These times are for a single processor,
if the problem could fit in memory.  Divide times by # of processors for
multi-processor runs.
		        *Run time to*	
Run size	Memory	t=10,	t=100,
Per side	MB	days	days
128		    90	  0.1	   1
256		   717	  0.9	   9
384		 2,419	  4.4	  44
512		 5,734	 13.9	 139
768		19,354	 70.3	 703
1024		45,875	222.0	2220

Conclusions:
128^3  is a good size to run tests and quickly explore parameter space.
256^3  will be my standard production size research run.
384^3  could be done for a short run (t=10) on one processor or a long run
       (t=100) on 3 or 4 processors. 
512^3  could reasonably run on two processors for a short run.  It 
       crashes on a single processor.
768^3  could run on Phillips (4 processor) for a short run (t=10)
       if you want to wait for 18 days for the results.
1024^3 could not fit on any of our shared memory machines, and would take
       too long if it did.