Opened 8 months ago

Last modified 7 months ago

#3490 new defect

Writing NetCDF ancillaries with ANTS does not scale

Reported by: vanniere Owned by: um_support
Component: UM Model Keywords: ANTS; scaling
Cc: Platform: JASMIN
UM Version: 11.0

Description

Hi CMS helpdesk

I have used ANTS to regrid aerosol clim ancillaries on the N2560 grid from the standard resolution ancillaries.
By default, ANTS always output a NetCDF version of the ancillary. It happens that writing a NetCDF with iris does not scale at such high resolution. For a single month and a single aerosol type, the job was killed after running over 48h wall clock time on SLURM. After commenting the line saving the NetCDF file in ANTS, the job was completed is less than 4h.

Since the UM format is saved after the NetCDF, I was wondering whether important modifications are made to iris cubes in the function saving NetCDF, which would prevent the UM to read ancillaries if I skipped this step.

Many thanks
Benoit

Change History (3)

comment:1 Changed 8 months ago by pmcguire

Hi Benoit:
I am not sure how much I can help.
But have you tried using iris2.2 or later? I was able to get massive speedup with supermean code in Python on JASMIN/SLURM from going from iris1.13iris2.2 and python2.7python3.7. The iris2.2 library uses dask instead of biggus. The iris2.2 library is the default with module load jaspy.
Patrick

comment:2 Changed 7 months ago by grenville

Hi Benoit

It would be strange if that were the case. How does the ancillary file look?

Grenville

comment:3 Changed 7 months ago by vanniere

Hi Patrick and Grenville

Thanks for your replies.

Patrick, I use python3.7, iris2.2, ants0.13.1.

Grenville, the ancil seems fine. I have been in contact with the Met Office. They generally recommend to output both NetCDF and UM format, but in that case not saving the NetCDF might the best immediate solution. They are looking into the problem of performance.

Thanks
Benoit

Note: See TracTickets for help on using tickets.