I am trying to run hofx simulations with high-resolution GEOS5 model outputs. It turns out that any FV3JEDI operation involving reading the model state will require 30-80 times the data volume in memory. For example, if the necessary model variables combined is less than 2 GB in disk space, a simple diffstate or hofx_nomodel operation will require 60-100 GB of memory to run successfully. This makes running jobs at native resolution drastically heavy and unfeasible. What is likely causing this problem?
Many thanks in advance!
Is this problem specific to GEOS or happens when reading GFS backgrounds too? You seem to be suggesting that the problem is in the read routine. Is that indeed where you believe the issue lies?
The simplest operation I had was diffstate with geos model state files, where the memory issue consistently (30-80 times magnification) shows up.
I haven’t tried anything substantial in the case of GFS yet. The closest comparison I had is between the diffstates_gfs and diffstates_geos in the tests of FV3-JEDI. The data required in these two cases are 105M for gfs and 13M for geos, The memory consumptions are 582M (~5.5x) for gfs and 431M (~33x) for geos.