Test                  /home/phaley/....    /gdata/projects/...

short PE run              0:08:09             5:48:36

Matlab large save         0:01:19             0:01:17
(all at once)

Matlab ASCII write        0:02:39             0:02:52
(PE granularity)

Matlab NetCDF write       0:11:16             0:16:00      
(PE granularity)
(sync'ing)

--------------------------------------------------------
tests on local disks

Matlab NetCDF write        0:11:44
(PE granularity)
(sync'ing) compute-8-7

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

Fortran writes (see test_write.f for code)

Write six (nx,nz) arrays for each (ny) slab (nt) times

  NX  NY NZ NT        /home/wgleslie/temp/     /gdata/WGL

 185 168 70  9             0:00:02              0:00:10
 185 168 70 18             0:00:10              0:00:16
1850 168 70 18             0:01:12              0:01:41
1850 168 70 36             0:02:27              0:03:20


Using an array of 185x168x70 and writing six arrays for each j slab nine times:

On mseas, writing to file in home area:

Mon Mar  7 09:47:17 EST 2011
Mon Mar  7 09:47:19 EST 2011

On mseas, writing to /gdata/WGL:

Mon Mar  7 09:58:42 EST 2011
Mon Mar  7 09:58:52 EST 2011

On compute-1-8, writing to /state/partition1/WGL:

Mon Mar  7 10:02:47 EST 2011
Mon Mar  7 10:02:50 EST 2011

On compute-9-10, writing to /state/partition1/WGL:

Mon Mar  7 10:04:18 EST 2011
Mon Mar  7 10:04:20 EST 2011

Note the 10 seconds when writing to /gdata versus 2-3 seconds when writing anywhere else. 

If I do 18 writes instead of nine, the number for the home area goes to seven seconds while
that on /gdata goes to seventeen (7 vs 17).  If I do this test 10 times in succession, I get
an average of ten seconds for the /home writes and sixteen seconds for the /gdata write.

If the nptsx is changed from 185 to 1850, the numbers increase to 1:12 for /mseas and 1:41
for the /gdata writes.

If 36 writes are done and nptsx is kept at 1850, the numbers further increase to 2:27 for
/mseas and 3:20 for /gdata.