Hi John, Dave, > The version NCEP is using is v3.6.3, so the issue here: > > http://www.unidata.ucar.edu/mailing_lists/archives/netcdfgroup/2011/msg00137.html > > would certainly be a concern; but perhaps not for WRF, since WRF always > writes all dimensions of a variable in order and in ascending index order > (referring to necessary condition #2 on the web page). Do you think that > means we're safe from this bug when using NOFILL? Yes. However, many other improvements and bug fixes have been made to netCDF, including netCDF-3, since version 3.6.3 was last released in June 2008, so I would recommend considering upgrading to a more current, supported version built with --disable-netcdf-4 to get the benefits of the enhancements and bug fixes. Using the --disable-netcdf-4 configure option means it's not necessary to build the HDF5 library. There have also been improvements and bug fixes to the netCDF-Fortran libraries since 3.6.3, which would require upgrading the netCDF-C library, because the netCDF-Fortran software is now a separately developed and maintained software package. Upgrading from 3.6.3 is something you could consider at any time in the future without worrying about backward compatibility, since more recent versions continue to support all the APIs and formats supported by 3.6.3, with added benefits for netCDF-3 users, including improved DAP access and easy transition to the netCDF-4 classic model format that supports compression and chunking. > Regarding the condition mentioned in your note, Russ, I there's no guarantee > that WRF will write variables in the order in which they're defined, though > it typically does. My interpretation of the statement in your note is that > there's a potential problem if the writing is interrupted (say, by program > failure?) but okay otherwise. I don't think we'd ever trust a WRF history > file that was written at the end of an abended run anyway, so we're probably > okay using NOFILL on that score too. Do you agree? Yes. > Related issue, Russ: Jim Abeles (IBM, CC'd here) also mentioned he was seeing > NetCDF read of a few MB for each large write operation. Can you say what is > causing that, and if it's necessary? When appending data to a record variable, it's necessary to read the last disk block in the record in order to keep the appended record contiguous with the previous record (modulo 4-byte alignment). Also, if the number of records is increased by a write, it's necessary to update that number in the header of a netCDF file, which requires reading and rewriting the count in the first disk block of the file. You might avoid the read of the last record disk block if you could arrange that each record contains a whole number of disk-blocks. But that would be non-portable, as the physical disk block size varies from platform to platform. However, it might be possible by padding each record with "data" of a size computed when the disk block size is known, at the beginning of a model run. > A clarification: the issue for NCEP's operational use of NetCDF in WRF isn't > runtime. WRF is writing through NetCDF asynchronously, so the speed of the > writes isn't usually a factor in model run time. The issue is contention > for limited system I/O bandwidth on the NCEP operational clusters. There are > a lot of I/O intensive jobs running at the same time and having the WRF jobs > generate 2x as much output traffic as needed has the effect of further > saturating the I/O system. Understood. > Dave, it is possible to add the NOFILL write mode to the WRFV3.6 netcdf > interface as an option that could be controlled at run time with a namelist > variable, then users would have a choice whether to use FILL or NOFILL. I > could do that. It would be up to the Developers Committee whether to make > the default FILL or NOFILL in WRF, but either way that would serve NCEP's > needs (assuming NOFILL is safe for their use-cases). It might be worthwhile to measure the difference in the context of operational use, to make sure using NOFILL really makes a significant difference in delaying or avoiding saturating the I/O system. --Russ > Thanks, > > John > > > -----Original Message----- > From: Unidata netCDF Support [mailto:address@hidden > Sent: Friday, February 14, 2014 6:26 AM > To: address@hidden > Cc: address@hidden; address@hidden; address@hidden > Subject: [netCDF #IJQ-160428]: NetCDF no-fill option > > Hi Dave, > > > I am one of the group that provides support to the WRF model. One of the > > most used data formats for the WRF model is NetCDF, largely because of the > > huge selection of post-processing tools that easily allows diagnostics and > > visualization. The WRF model is used in production at NOAA, and as such > > the operations staff are always quite sensitive to the amount of time any > > model takes. A simple method to reduce the wall-clock time is taken > > seriously. > > > > The NOAA developers have asked a couple of questions about the FILL / > > NOFILL option (here we reference this UNIDATA page): > > http://www.unidata.ucar.edu/software/netcdf/docs/netcdf-c/nc_005fset_005ffill.html > > > > In our model history volumes, if we only open the files for write, then the > > NOFILL option seems to be (not only) an OK choice, but perhaps one that > > provides a timing reduction. Our tests indicate that a write with the > > NOFILL option is a 1x write, compared to a 2x write + 1x read when doing > > the FILL option. Would you please verify this for us? We think that we > > would like to set the NOFILL option as the default option in WRF, to > > benefit our entire community. Would you please comment on our possible > > plans in this regard to help us know if this is the right move. All of our > > data has TIME as the unlimited dimension. Other than the HDF5 compression > > capabilities from NETCDF4, we are a pretty vanilla user group of NetCDF. > > I agree that setting the NOFILL option will save time in writing your model > history volumes, > and would be a reasonable default. The issue is whether the performance > benefit outweighs > any problems that might occur in determining what data was written and what > data are > essentially garbage values in case you write the history data variables in a > different order > used to define the variables in creating the file. For example, if you write > the last record > variable first and the writing is interrupted, then the values of the > previous record variables > in that record could be arbitrary. But if you log the creation of history > volumes so you know > when the writing is complete and buffers are flushed, either by a close or > sync call, that seems > like it shouldn't be a problem. > > One other issue is a potentially serious bug related to use of nofill mode > and a specific pattern > of writing data in versions of netCDF before version 4.1.3 (released in June > 2011): > > [netcdfgroup] Important: potential file corruption using NOFILL mode > http://www.unidata.ucar.edu/mailing_lists/archives/netcdfgroup/2011/msg00137.html > > If you're sure that WRF users use version 4.1.3 or greater of the netCDF C > library, or that > WRF doesn't meet the conditions described above that trigger the bug, then I > think it's safe. > > There was also a less serious bug with use of fill mode in netCDF versions > before 4.2, > released in May 2012: > > Fixed turning off fill values in HDF5 layers when NOFILL mode is set in > netCDF-4 API > https://bugtracking.unidata.ucar.edu/browse/NCF-151 > > > Secondly, we are interested in knowing about the statement: > > The use of this feature may not be available (or even needed) in future > > releases. Programmers are cautioned against heavy reliance upon this > > feature. > > If we get timing reductions, we would like to rely on the continuing > > availability of this feature. > > I think that statement was overly cautious, because we didn't know whether we > could support > nofill mode in future versions. Now we're committed to backwards > compatibility for such > features, and I don't anticipate that support for nofill mode would be > dropped. The HDF5 > software layer used by netCDF-4 also supports fill values and > > > Thanks in advance for your time and explanations. > > You're welcome, I hope this helps. > > --Russ > > > Dave > > > > > > David Gill > > Software Engineer > > address@hidden > > office: (303) 497 8162 > > fax: (303) 497-8171 > > Mailing address: > > 3450 Mitchell Lane > > Boulder, CO 80301 > > > > > > > > > Russ Rew UCAR Unidata Program > address@hidden http://www.unidata.ucar.edu > > > > Ticket Details > =================== > Ticket ID: IJQ-160428 > Department: Support netCDF > Priority: Normal > Status: Closed > > Russ Rew UCAR Unidata Program address@hidden http://www.unidata.ucar.edu Ticket Details =================== Ticket ID: IJQ-160428 Department: Support netCDF Priority: Normal Status: Closed
NOTE: All email exchanges with Unidata User Support are recorded in the Unidata inquiry tracking system and then made publicly available through the web. If you do not want to have your interactions made available in this way, you must let us know in each email you send to us.