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

Re: netcdf i/o on ibm vs cray

>To: address@hidden
>From: Ben Foster <address@hidden>
>Subject: Re: 20000207: netcdf i/o on ibm vs cray
>Organization: HAO
>Keywords: slow netCDF I/O on Cray

Hi Ben,

> I have written a netcdf file with time as the unlimited
> dim, with 7482 records. The time dimension is called "day"
> and there is an int coord var called day(day) (yyddd). 
> 
> When I do an ncdump -c on the file on the ibm (e.g. blackforest), 
> it prints the 7482 integers in a few seconds, but on the Cray
> (chipeta) it takes several minutes. I don't think this is
> happening because of load on the machines. Is the i/o really
> that much slower on the Crays than the IBM?
> 
> FYI, the file is on mss /TGCM/data/ncep_79001-99177.nc
> (about 161 MB). Here is the ncdump with the first 20 days:

First, sorry it's taken so long to reply; I've been going to a lot of
all-day meetings recently.

The default configuration for use on the Cray uses an FFIO
specification that favors sequential access, but you can probably do
much better by playing with the NETCDF_FFIOSPEC environment variable.
For some possible specifications for various I/O scenarios, see:

  http://www.unidata.ucar.edu/packages/netcdf/guidec/guidec-14.html#HEADING14-0

or try setting NETCDF_FFIOSPEC to "cachea:8:256:0" before running
ncdump to see if it makes a difference.  Some users have reported huge
performance differences on Crays by changing this FFIO specification.
See, for example:

  http://www.unidata.ucar.edu/glimpse/netcdf/3261

--Russ