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

Re: transition from crays to ibm

>To: address@hidden
>From: Ben Foster <address@hidden>
>Subject: Re: 20020905: transition from crays to ibm 
>Organization: NCAR/HAO
>Keywords: external vs. internal data types

Ben and Rashid,

> On the IBM's we use -qrealsize=8 (8-byte reals, allowing default
> 4-byte integers) and link the r4i4 netcdf library. When we are
> ready to open a history file, we tell netcdf to define fields going 
> onto the history as "doubles", i.e., 8-byte reals:
> 
> istat = nf_def_var(ncid,f%short_name,NF_DOUBLE,4,idimids,idvar)
> 
> Then when it comes time to write values to the file, we give it
> 8-byte reals (which were declared "real" in the model):
> 
> istat = nf_put_vara_double(ncid,idvar,start_4d,count_4d,f3d)
> 
> The r4i4 netcdf library can do i/o on either 4-byte or 8-byte 
> reals. The fact that the library was compiled with -r4 does not 
> affect this ability (Russ: is this strictly true?). ...

Yes, the numeric type of internal data that is written to a netCDF
file (int, real, double) is independent from the type of the netCDF
variable on disk to which it is written (NF_BYTE, NF_SHORT, NF_INT,
NF_REAL, NF_DOUBLE).  Conversion happens during the write.  If you
try to store 8-byte internal reals into a 4-byte external NF_REAL
variable, you will lose precision.  The only thing the "-r4" the
library was compiled with affects is whether internal REAL variables
are 4-byte or 8-byte.

> ... When we do
> post-processing, we still link the r4i4 library, compile with -r4, 
> and read the doubles into the 4-byte reals with nf_get_vara_real(...). 
> For vis, we accept the loss of precision from 8-byte to 4-byte.
> This is probably what your IDL code is doing also.
> 
> I'm assuming this is what most large models are doing (e.g., ccm, 
> csm, etc), so I'm surprised the consultants suggested that you 
> rebuild the netcdf library yourself! (I am cc'ing the ncar consultants 
> and Russ Rew).
> 
> --Ben 
> 
> 
> >Date: Thu, 05 Sep 2002 11:06:33 -0600
> >From: "Rashid Akmaev" <address@hidden>
> >X-Accept-Language: en
> >MIME-Version: 1.0
> >To: Ben Foster <address@hidden>
> >Subject: Re: transition from crays to ibm
> >Content-Transfer-Encoding: 7bit
> >
> >Hi Ben,
> >
> >I of course waited till the last moment to move from chipeta to bf.  At
> >least I was able to convert most of my model input/output to netCDF, as
> >you recommended.  It worked fine, for example, with idl here.  Now, I've
> >just found that the corresponding netcdf library on bf is only available
> >in r4i4 format.  Naturally, I'd like to run everything in the r8 size to
> >immitate the CRAYs (not sure about the integers yet) but there is no r8
> >netcdf library.  I tried the r4i4 version and not everything works as
> >expected with r8, and my understandning is that it's not recommended to
> >mix sizes.  Consulting says that I should rebuild the library myself for
> >the size I want.  How did you go about this?