CABLE benchmarking

Repository to benchmark a user branch (with personal changes) against the head of the trunk. The results should be attached with all new tickets.

The code will: (i) check out; (ii) build; and (iii) run both the head of the trunk and the user's personal branch (or integration branch) across N standard science configurations.

NB. the code is flexible enough that "trunk" could be any branch, allowing the user to compare across personal branches.

To get started:

$ git clone https://github.com/mdekauwe/CABLE_benchmarking.git

Setup integration branch

This ought to be a one time thing, but just to remind myself how we got this...

$ python scripts/setup_integration_branch.py

Multi-site comparison

Runs both cable executables (trunk and the user's branch/integration branch) across a suite of FLUXNET sites and plots seasonal cycles. Whilst these plots are generated locally, the results of each repository should be uploaded to Gab's benchmarking website for further benchmarking and evaluation.

To run the multi-site comparison, please update the relevant entries (paths, library locations, etc) within user_options.py

This should be pretty self-evident, the expectation is that the user will only change entries between:

#------------- User set stuff ------------- #
user = "XXX579"

...
# ------------------------------------------- #

In many cases we recommend you don't change the defaults, e.g. (however things should be robust to you doing as you please)

src_dir = "src"
run_dir = "runs"
log_dir = "logs"
plot_dir = "plots"

To pass different science configurations, the code expects the details to be passed as a dictonary. All options passed are added to the cable namelist file, so it is pretty flexible.

There are a couple of potential gotchas:

To pass a string flag, you need a double set of quotation marks (i.e. double and single) as shown below. If you encounter a problem this is likely to be the issue ...

sci = {
    "cable_user%GS_SWITCH": "'medlyn'",
}

This isn't an issue for other flags, e.g.

sci = {
    "output%restart": ".FALSE.",
    "fixedCO2": "400.0",
}

On the NCI, if you get an error about your python version:

"The use of the #!/usr/bin/env python interpreter line in python scripts..."

Then try:

module load python3-as-python

Single-site, multi-sites and all-sites ...

To run a single site, you just need to set which site you wish to run in the python list variable, e.g.

met_subset = ['TumbaFluxnet.1.4_met.nc']

This structure is very flexible, so if you want to run three sites, you just expand the entry:

met_subset = ['TumbaFluxnet.1.4_met.nc', 'HarvardFluxnet.1.4_met.nc','HowardFluxnet.1.4_met.nc']

And to simply run all the met sites in a directory, leave the list empty:

met_subset = []

Finally, if you are running more than a single-site, there are two MPI flags you should consider setting to speed up things:

mpi = True
num_cores = 4 # set to a number, if None it will use all cores...!

After updating user_options.py, to run the code

$ ./run_site_comparison.py

If you're on the NCI and wish to submit a qsub script instead:

$ ./initialise_qsub_job.py
$ qsub benchmark_cable_qsub.sh

There are two steps here as the NCI nodes don't have internet access, so we need to check out and build CABLE first.

If you've already built the src once and just want to run things again:

$ ./initialise_qsub_job.py -s

will generate the qsub script and

$ ./run_site_comparison.py -s

will run a local benchmarking without trying to download and rebuild src code

If you want to make some quick local benchmark plots:

$ ./make_seasonal_plots.py

Global comparison

Coming soon ...

Code dependencies

The code has been written such that it has very few dependancies to ease personal set up. Nevertheless, it does depend on a few fairly standard python libraries:

• numpy
• pandas
• xarray

All of which can be easily built using anaconda.

To install on raijin in your personal space:

Download the linux anaconda file. Then create an environment called "science" (or whatever sensible name springs to mind).

$ conda create --name science python=3

then

$ source activate science

then

$ conda install docopt xarray matplotlib pandas scipy numpy

If you're working locally on a mac or linux machine, you could as easily use your favourite package manager (e.g. macports, apt-get, etc).

Contacts

• Martin De Kauwe.
• Gab Abramowitz.