On the use of reduced precision

[martinjuckes]

Matt
Not directly related to FORTRAN types; I wonder if there is capacity here for considering stored precision of the data we have. There are studies considering the benefits of lossy compression, e.g. Zender (doi:10.5194/gmd-2016-63), Baker et al (doi:10.5194/gmd-9-4381-2016), , and being able to specify variables are held with limited precision enough could lead to significant savings in storage costs.

Martin N Juckes
I can put in something to specify what the precision is -- but the biggest problem is trying to decide on any reduction. There are some issues around diagnosing what is going on in the model. E.g. data may have a lot of noise in which has no physical meaning but gives clues about what is going on in the model ... perhaps we don't care about that sort of data in CMIP? Getting a lossy compression algorithm into the NetCDF library would make it a lot easier.