[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Unidata Support: 951212: netCDF-2.4-beta5 test results from GFDL

Hi Jeff,

Here's a note from someone who tested your netCDF Cray optimizations, and is
mostly quite happy with them.

Do you remember anything about NCFILL vs. NCNOFILL, and whether the
optimizations were just for NCNOFILL?

--Russ

------- Forwarded Message

>From: address@hidden (John Sheldon)
>Organization: GFDL
>Keywords: 199512121934.AA29804 netCDF CRAY

Hi (Steve?)-

Just thought I'd let you know that I recently did some simple timing
tests on our C90 with version 2.4, and compared the times to writes
using version 2.3 and to FORTRAN writes using both native Cray and IEEE
formats.  In my tests I wrote out, x-y slab by x-y slab, an array which
was 360x180x20x3.  The native Cray writes took about .3 user user-CP
seconds, about .2 system-CP sec, and between 20 and 52 wall-clock
seconds (the lesser value was for a direct-access version).  The IEEE
writes took about .45 user-CP sec and ~.4 system-CP sec, but about 3 to
3.5 minutes wall clock (I can't explain this!).

The netCDF 2.3 write, with NCNOFILL, took ~21 CP sec, .5-.8 system-CP
sec, and about 30-40 sec wall-clock (I ran it a few times).  The netCDF
2.4 write took .62 user-CP sec, .25-.4 system-CP sec, and only 6-8 sec
wall clock!! Way to go!!  (How did you reduce the wall clock time like
that??)

The one mystery that remains occurs when using pre-filling (NCFILL).
This is fairly serious, since I urge all our local users to employ
pre-filling.  For netCDF 2.3, it took ~39 user-CP sec (only ~1
system-CP sec) and 1:26 wall-clock, about double the NCNOFILL time and
a not-too-unexpected result.  But for netCDF 2.4, it still took 38 CP
sec (plus 1.5-4 system-CP sec) and 46 sec wall-clock -- hardly any
change!  Is there a piece of code not optimized somewhere?

Thanks for any info you can provide.  Hope my results are helpful.

John Sheldon
address@hidden

------- End of Forwarded Message