pNetCDF and WRF

[John Michalakes]

Hi Rajeev and list

Here are some preliminary results from the NCAR IBM Power5 system
(bluevista) comparing parallel NetCDF vs regular NetCDF output in the WRF
model. This is for one of the standard WRF benchmark cases (12km CONUS; see
http://www.mmm.ucar.edu/wrf/WG2/bench) where each write is about 250 MB.  I
timed writes to a new NetCDF file and writes to an already existing NetCDF
file separately.

WRF has an ansynchronous output capability that collects up the data onto a
bank of dedicated output tasks which then write the data while the compute
tasks go forward. Performance of regular NetCDF output with this async
output mode is shown on one of the other tabs of the attached spreadsheet.
For this series of runs, the async output generally does better than
parallel NetCDF, but has to collect up all the data on an output server
task. This collection seriously limits the size problem we can run,
especially on very large numbers of BG processors, and was the main
motivation for going to pNetCDF.  Next step will be to try some runs on the
Blue Gene system here at NCAR.

If anyone's intersted in obtaining this new version of WRF that uses pNetCDF
for your own testing, please let me know.

Regards,

John

> -----Original Message-----
> From: owner-parallel-netcdf at mcs.anl.gov
> [mailto:owner-parallel-netcdf at mcs.anl.gov]On Behalf Of Rajeev Thakur
> Sent: Sunday, October 29, 2006 8:12 PM
> To: john at michalakes.us; parallel-netcdf at mcs.anl.gov
> Subject: RE: pNetCDF and WRF
>
>
> Hi John,
>         That is good news. Did you have a chance to see what kind of
> performance you get with PnetCDF versus regular netCDF?
>
> Rajeev
>
> > -----Original Message-----
> > From: owner-parallel-netcdf at mcs.anl.gov
> > [mailto:owner-parallel-netcdf at mcs.anl.gov] On Behalf Of John
> > Michalakes
> > Sent: Sunday, October 29, 2006 1:59 AM
> > To: parallel-netcdf at mcs.anl.gov
> > Subject: pNetCDF and WRF
> >
> > An update and a question:
> >
> > Update: Parallel NetCDF is now successfully integrated as an
> > I/O option into
> > the Weather Research and Forecast Model. Thanks for your help
> > answering
> > questions these last couple of weeks. This new capability
> > will be included
> > in the upcoming WRF 2.2 release in December 2006. pNetCDF
> > will be supported
> > but not required in WRFv2.2 (it will be possible to build and run WRF
> > without it). We will not distribute pNetCDF with WRF; rather,
> > users will be
> > directed to install pNetCDF from the ANL distribution.
> >
> > Question: what is the status of pNetCDF on Blue Gene, please?
> > Any caveats or
> > other special information?
> >
> > Thanks again,
> >
> > John
> >
> > -----------------------------------------------------------
> > John Michalakes, Software Engineer        michalak at ucar.edu
> > NCAR, MMM Division                   voice: +1 303 497 8199
> > 3450 Mitchell Lane                     fax: +1 303 497 8181
> > Boulder, Colorado 80301 U.S.A.        cell: +1 720 209 2320
> >         http://www.mmm.ucar.edu/individual/michalakes
> > -----------------------------------------------------------
> >
> >
>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: PNetCDF 12km CONUS BV.xls
Type: application/vnd.ms-excel
Size: 26112 bytes
Desc: not available
URL: <http://lists.mcs.anl.gov/pipermail/parallel-netcdf/attachments/20061030/1cc8d5ef/attachment.xls>