[Go-essp-tech] CMOR and cell_measures issues

Mon Nov 1 11:09:30 MDT 2010

Hello Karl,

thanks for this reply.  Putting aside the issue of whether this is
really ext_cell_measures or cell_measures then I think, given the
resources we have locally, we have to make a choice of correctness vs
completeness.  The reason we are tempted to turn off ext_cell_measures
is it is the least effort way we can see of submitting data that is
correct.  I think you are suggesting going for completness - even if we
risk submitting some data with ext_cell_measures that is incorrect.
Obviously this is *my* interpretation of what you are saying.  Yes we
can go for both correctness and completeness, but this will take us some
effort - we need an exta component in our system that can recognise
which cell areas to assign to which variables (with minimum error) - and
we (like everyone) have lots of demands on our effort at the moment -
and we have to make judgements about where to prioritise.  (This isn't
supposed to be a sob story - just trying to explain why we are
tempted...)

Would you recommend 'completeness' over 'correctness' - have I
interpreted you correctly?  What are the options for correcting
incorrect meta-data once data is ingested into ESG?

Jamie

________________________________

	From: Karl Taylor [mailto:taylor13 at llnl.gov] 
	Sent: 29 October 2010 21:36
	To: Kettleborough, Jamie
	Cc: Bentley, Philip; V. Balaji; martin.juckes at stfc.ac.uk;
go-essp-tech at ucar.edu; cmor at lists.llnl.gov; Kyle.Olivo at noaa.gov;
Doutriaux, Charles
	Subject: Re: [Go-essp-tech] CMOR and cell_measures issues

	Dear Jamie and Charles (a couple of questions for you),

		Hello Karl,

		I think the recommended way to 'turn off'
ext_cell_measures is to make a call to
cmor.set_variable_attribute(varid, 'ext_cell_measures', '').  Is that
right?  We are very tempted to do this for all variables - so basically
overriding the MIP tables.  How big a problem do you think this will be
for data users - our grid is pretty straight forward and users can
calculate cell_areas from the latitudes.

	Yes, if the cell areas stored in areacella are not appropriate
for a particular field, and the requested output tables say that
ext_cell_measure includes areacella, then you should call the set
attribute function to reset ext_cell_measures="".  Isn't that right
Charles?

	Why are you tempted to turn off the ext_cell_measures for all
variables?  Then your output won't conform to the CMIP5 requirements.

	In the latest CMOR tables, I have removed ext_cell_measures from
all the variables that we don't expect always to be on the standard mesh
(i.e., on the grid for which areacella is correct).  This includes
velocities and transports and closely related fields, which are
sometimes staggered relative to areacella.  I would still be interested
in hearing a clear explanation for why there are additional fields
carried on a completely different grid. 

	If users must compute the cell areas for only your grid, and for
all others they simply read the areacella field in, then you are
creating a special case that is completely unnecessary.

		That aside, doesn't the approach of providing
alternative grid areas need more discussion?

		  1. how should we produce these.  The most natural
approach I can think of is to modify the fx MIP tables to add in
areacellb (or whatever we choose to call it) and then output through
CMOR - this will maximise the chance of consistency between different
grid area files for any one model. 

		  2. how should we reference these additional areas from
a variable.? I could call cmor.set_variable_attribute(varid,
'ext_cell_measures', 'areacellb') - but in the tests I've done on CMOR
2.4 this only does half the job: it puts the appropriate
ext_call_measures attribute into the file, but does nothing with
associatedFiles.

	I don't think it is a high priority to standardize this
immediately.  We will want CMOR to place the fields in the subdirectory
fx, so I need to check with Charles whether this requires the variable
to appear in table fx.  If not, I would probably build an entirely new
table similar to fx, but with only the additional variables.  This way
you won't have to modify your table if a new fx table comes out.  As for
referencing these additional area variables, I think if you include
area: <area_name> in the ext_cell_measures attribute, then if CMOR isn't
already doing this, Charles can modify construction of associated_files
to include something following the template "<area_name>:
<area_name>_fx_IPSL-CM5_historical_r0i0p0.nc"   What do you think,
Charles?

		Clearly these may have been things you were going to
cover - but ran out of time to, in which case sorry.

		I think another scenario that still needs some thought
is one where a data provider has submitted data and published it in ESG.
They then realise they made a mistake - they should have turned
ext_cell_measures off, but didn't (or visa-versa). What happens in this
case?  (We have kind of done this in that we have send data with
incorrect cell_measures to the BADC - but have caught the issue before
ingestion into ESG  - I don't believe we will always be this lucky).
You'll probably see through why I'm asking this question about meta-data
updates again now, so I may as well be explicit... If we choose to turn
off ext_cell_measures for all our diagnostics on this initial submission
- what are our options for recovering from this if we later found the
decision to submit without ext_cell_measures was making our data hard to
use?

	Please don't turn off ext_cell_measures (in general).   I think
you could easily write a script to remove the cell_measures attribute
using netCDF tools, but adding it would require rewriting the entire
file.

	Best regards,
	Karl

		Jamie

________________________________

			From: Karl Taylor [mailto:taylor13 at llnl.gov] 
			Sent: 29 October 2010 02:15
			To: Bentley, Philip
			Cc: V. Balaji; martin.juckes at stfc.ac.uk;
go-essp-tech at ucar.edu; cmor at lists.llnl.gov; Kyle.Olivo at noaa.gov;
Doutriaux, Charles; Kettleborough, Jamie
			Subject: Re: [Go-essp-tech] CMOR and
cell_measures issues

			Dear all,

			I meant to try to address all the stuff in this
discussion, but won't have time today.  This email is just to say that I
think we should insist that the cell_area files (areacella and
areacello) be placed in the archive, even if there are also gridspec
files.   The ext_cell_measures attribute should also be included for
fields that are on the "standard" grid (i.e., the one with the cell
areas stored in areacella or areacello).  If there are other fields for
which the standard areas are inappropriate and where your scientists
think it is important to provide cell areas, then I recommend that you
create specially named variables and place them in the "fx"
subdirectories.   For variables not on the "standard" grid (i.e., the
grid of areacella or areacello), you should "turn off" the
ext_cell_measures attribute.

			I don't expect most groups to produce gridspec
files, so most analysts will be looking for areas in the areacella and
areacello variables, not the gridspec files.  This is why you should
write the areacella and areacello files even if you also write the
gridspec files.

			Also, could you please explain why you prefer
not to duplicate the "fx" fields in each experiment's directory tree. 

			Best regards,
			Karl

			On 10/25/10 7:12 AM, Bentley, Philip wrote: 

				Hi Balaji,

				Phil, I'm very impressed that Had will
have gridspec files, 
				is this a done deal? I've been so
pessimistic about this that 
				I was wondering if even we should do one
ourselves.

				Nope, not a done deal yet :-(

				In line with the CMIP5 expt design doc,
we don't really need to provide
				gridspec files since all our model
output is on either regular or
				uniform grids (i.e. simple cartesian
product of lat & long).

				However, this whole cell_measures
business prompted me to revisit the
				gridspec tools and output, which
reminded me that the gridspec netcdf
				files include a cell area variable.
Which in turn means we wouldn't need
				to provide a separate file (or files)
for cell areas. Hence we could
				drop the ext_cell_measures attribute
from our CMIP5 output files.

				Using the gridspec tools may be a quick
and easy way for us to provide
				cell area info if we need to.

				Caveat: from a quick glance it looks
like the netcdf files produced by
				the gridspec tools are not CF compliant.
Is this is an issue? Presumably
				it is if we want all the data in the
CMIP5 archive to be CF compliant.
				(NB: it could be I'm not running with
the very latest version of the
				tools - but I couldn't see a more recent
version on the gfdl web site).

				You know of course that gridspec says
you can supply

				gridspec_fx_HadGEM2-ES_atm_pgrid.nc
				gridspec_fx_HadGEM2-ES_atm_ugrid.nc
				gridspec_fx_HadGEM2-ES_atm_vgrid.nc
				gridspec_fx_HadGEM2-ES_atm_uvgrid.nc

				as one single supergrid...

				If I could figure out how to output all
7 or 8 atm/ocn (sub-)grids to a
				single netcdf file I would, but the
available documentation (e.g. for
				make_hgrid) isn't clear on this point.
Sorry, that's probably just me
				being dumb! But if there is updated
documentation then please point me
				to it. If necessary I could concatenate
variables afterwards using NCO
				tools.

				Right now I'm trying to figure out how
to create a gridspec file for our
				HadGEM2 ocean model, which uses a
stretched (i.e. tartan/plaid) grid:
				longitudes are evenly spaced, latitudes
vary from 1 deg to 1/3 deg.
				(Looks like I need to use the
--my_grid_file option to supply the
				lat/long coords).

				But if you're doing gridspec at all, I
will concede this 
				point:-). Let's both do these separate
gridspecs for now.

				Works for me.

				I think we're suffering from
'early-adopter syndrome' :-/

				Cheers,
				Phil

				Bentley, Philip writes:

				Hi Karl,

				A somewhat belated follow-up question in
connection with 

				this proposal 

				(and with some slight overlap with
Jamie's email which 

				crossed on the 

				ether)...

				As things stand the files named in the
'associated_files' attribute 
				appear thus (using our RCP 4.5
simulation as an example):

				"... gridspecFile:
gridspec_fx_HadGEM2-ES_rcp45_r0i0p0.nc areacella:
				areacella_fx_HadGEM2-ES_rcp45_r0i0p0.nc"

				Are the <expt_id>_<rip> parts (i.e.
'rcp45_r0i0p0.nc' ) actually 
				required? AFAIK, our gridspec/cellarea
files will not 

				change from one 

				simulation to the next using the same
model (HadGEM2-ES in 

				this case).

				Since, like most centers, we will be
running large numbers of 
				simulations using the same model, it
looks like we would need to 
				create numerous duplicates of the
gridspec/cellarea files - 

				or lots of 

				symlinks
				- in order to for these references to
make sense. Unless you are 
				planning to manage that on our behalf
somehow...?

				I think our 4 gridspec files for the
HadGEM2 atm grids are 

				likely to 

				be called something like...

				gridspec_fx_HadGEM2-ES_atm_pgrid.nc
				gridspec_fx_HadGEM2-ES_atm_ugrid.nc
				gridspec_fx_HadGEM2-ES_atm_vgrid.nc
				gridspec_fx_HadGEM2-ES_atm_uvgrid.nc

				So without any simulation-specific info.
(There would also be files 
				for the ocean grids)

				As it happens the gridspec files contain
grid cell areas, 

				so I'm now 

				wondering if we'd even supply both?

				I'd be interested to hear your thoughts
on this. I may be 
				mis-understanding something/everything
:-)

				Regards
				Phil

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.ucar.edu/pipermail/go-essp-tech/attachments/20101101/00db6284/attachment-0001.html 