paradigm

Requirements

All command suggestions below assume a Linux operating system with makefile installed.

Setup a conda environment

The easiest way to set up all the needed tools is to create an Anaconda environment from the given environment.yml. Just type:

make conda

Then, before running all the commands below make sure you activate the environment

source activate dicedps

If working with a Modules package, make sure you are in a clean state:

module purge
module load anaconda3/4.3.0 cmake/3.8.2 gcc/5.3.1 netcdf/4.4.0 openmpi/1.10.1 openjdk/1.8.0
export PYTHONPATH=

Install BRICK dependencies

Type:

make brick-install

to install the missing R packages and compile the Fortran submodules needed for BRICK.

Generate BRICK calibration ensembles

Create an output directory

mkdir -p output/brick

Test the BRICK calibration machinery with:

make brick-calib-cauchy NCHAIN=2 NITER=10000

This should create a file brick_mcmc_fgiss_TgissOgour_scauchy_t18802011_z18801900_o4_h50_n10000.rds under output/brick containing two Markov chains with 1e4 samples each.

Change the definition of LONGRUN in makefile_common to match your setup for running computationally expensive make targets. In my case I have a bash function qmake.sh that creates a .pbs file and submit a job to a cluster to make the given target with the required resources.

By default, NITER (number of MCMC samples) is 20e6, and NCHAIN (number of MCMC parallel chains) is 10. To reduce the resources needed, decrease NITER and/or NCHAIN in pkgs/brick/makefile. Then run from the main directory:

make brick-calib-all

Diagnostics and the final chains are created with:

make brick-diag-all

Your output directory should look like this:

output/
└── [4.0K]  brick
    ├── [826K]  brick_mcmc_fgiss_TgissOgour_scauchy_t18802011_z18801900_o4_h50_n10000.rds
    ├── [9.4G]  brick_mcmc_fgiss_TgissOgour_scauchy_t18802011_z18801900_o4_h50_n20000000.rds
    ├── [ 17M]  brick_mcmc_fgiss_TgissOgour_scauchy_t18802011_z18801900_o4_h50_n20000000_t1000_b5-chains.rds
    ├── [6.3K]  brick_mcmc_fgiss_TgissOgour_scauchy_t18802011_z18801900_o4_h50_n20000000_t1000_b5-hd.rds
    ├── [ 16M]  brick_mcmc_fgiss_TgissOgour_scauchy_t18802011_z18801900_o4_h50_n20000000_t1000_b5.nc
    ├── [832K]  brick_mcmc_fgiss_TgissOgour_schylek_t18802011_z18801900_o4_h50_n10000.rds
    ├── [9.4G]  brick_mcmc_fgiss_TgissOgour_schylek_t18802011_z18801900_o4_h50_n20000000.rds
    ├── [ 19M]  brick_mcmc_fgiss_TgissOgour_schylek_t18802011_z18801900_o4_h50_n20000000_t1000_b5-chains.rds
    ├── [6.3K]  brick_mcmc_fgiss_TgissOgour_schylek_t18802011_z18801900_o4_h50_n20000000_t1000_b5-hd.rds
    ├── [ 18M]  brick_mcmc_fgiss_TgissOgour_schylek_t18802011_z18801900_o4_h50_n20000000_t1000_b5.nc
    ├── [829K]  brick_mcmc_fgiss_TgissOgour_spaleosens_t18802011_z18801900_o4_h50_n10000.rds
    ├── [9.4G]  brick_mcmc_fgiss_TgissOgour_spaleosens_t18802011_z18801900_o4_h50_n20000000.rds
    ├── [ 15M]  brick_mcmc_fgiss_TgissOgour_spaleosens_t18802011_z18801900_o4_h50_n20000000_t1000_b5-chains.rds
    ├── [6.4K]  brick_mcmc_fgiss_TgissOgour_spaleosens_t18802011_z18801900_o4_h50_n20000000_t1000_b5-hd.rds
    └── [ 14M]  brick_mcmc_fgiss_TgissOgour_spaleosens_t18802011_z18801900_o4_h50_n20000000_t1000_b5.nc

where the *_t1000_b5{.nc,-chains.rds} files contain the thinned chains, and the *-hd.rds files contain diagnostic information.

Checking the output will give you information on the number of samples remaining.

Install BORG

Make sure you have rights to clone the serial-borg-moea repository. Then:

make borg

If compilation succeeds, you should have the following files among others:

pkgs/borg/build
|-- bin
|   |-- borg
|   |-- dtlz2_advanced
|   |-- dtlz2_ms
|   `-- dtlz2_serial
`-- lib
    |-- libborg.so
    `-- libborgms.so

Check that both serial and parallel versions work:

cd pkgs/borg/build/bin
./dtlz2_serial
# (make sure last number shown is 0.746296991175737445268)
mpirun -np 10 ./dtlz2_ms
# (stop with Ctrl-C and check runtime_0.txt)

Install Python requirements

From the Anaconda environment, type:

pip install -r requirements.txt

Check that the Python bindings with Borg work:

python pkgs/borg4platypus/examples/simple_borg.py

Extrapolate forcings

Historical forcings data stops at 2011, an automatic ARIMA is used to estimate missing forcings up to 2015.

make forcings

Run optimization

Create an output dir:

mkdir output/dicedps

Check that optimization works on a simple instance:

make opt-test

Then check your cluster setup:

make opt-mini

If everything works, proceed with the full-scale optimization

make opt-full

This should give you the following runtime files:

output/dicedps/
├── u1w1000doeclim_mrbfXdX41_i1p400_nfe5000000_objv2_cnone_s1_seed0001_runtime.csv
├── u1w1000doeclim_mrbfXdX41_i1p400_nfe5000000_objv2_cnone_s2_seed0002_runtime.csv
├── u1w1000doeclim_mrbfXdX41_i1p400_nfe5000000_objv2_cnone_s3_seed0003_runtime.csv
├── u1w1000doeclim_mrbfXdX41_i1p400_nfe5000000_objv2_cnone_s4_seed0004_runtime.csv
├── u1w1000doeclim_mrbfXdX41_i1p400_nfe5000000_objv2_cnone_s5_seed0005_runtime.csv
├── u1w1000doeclim_mtime_i1p400_nfe5000000_objv2_cinertmax_s1_seed0001_runtime.csv
├── u1w1000doeclim_mtime_i1p400_nfe5000000_objv2_cinertmax_s2_seed0002_runtime.csv
├── u1w1000doeclim_mtime_i1p400_nfe5000000_objv2_cinertmax_s3_seed0003_runtime.csv
├── u1w1000doeclim_mtime_i1p400_nfe5000000_objv2_cinertmax_s4_seed0004_runtime.csv
└── u1w1000doeclim_mtime_i1p400_nfe5000000_objv2_cinertmax_s5_seed0005_runtime.csv

with Borg iterations.

Merge Pareto fronts

We’ll use PyPy for speeding things up. Make sure requirements are available.

pypy3 -m ensurepip
pypy3 -m pip install tqdm

Check Pareto merging with:

make par-test

Run the actual merge with:

make par-merged

This should give you the following merged Pareto files:

output/dicedps/
├── u1w1000doeclim_mtime_i1p400_nfe5000000_objv2_cinertmax_s0_seed0000_merged.csv
├── u1w1000doeclim_mrbfXdX41_i1p400_nfe5000000_objv2_cnone_s0_seed0000_merged.csv

Check BORG runtime metrics

Install MOEAframework and other required tools.

make met-setup

Check that the setup works

make met-test

This should produce a file output/dicedps/u1w1000doeclim_mtime2_i1p400_nfe5000000_objv2_cnone_s1_seed0001_metrics.csv with the following metrics table:

#NFE ElapsedTime SBX DE PCX SPX UNDX UM Improvements Restarts +hypervolume
250000 2148.925116 0.182977 0.000784468 0.81565 0.000196117 0.000196117 0.000196117 34853 0 0.5646850294
500000 4292.769322 0.0343127 0.000328875 0.96503 0.000109625 0.000109625 0.000109625 62819 0 0.5666289431
750000 6439.791670 0.0143852 9.04732e-05 0.985253 9.04732e-05 9.04732e-05 9.04732e-05 84237 0 0.5675188842
[...]

Run

make met-all

to compute the metrics for all runtime files.

Reevaluate Pareto fronts

Check the rerun script with:

make rerun-test

Rerun the whole thing with:

make rerun_v3

Regenerate manuscript figures

Run the fig_*.py scripts under pkgs/dicedps/dicedps/plot/. Make sure to run the *_data.py first.

License

All the code in this repository, unless specified otherwise, is released under the GNU General Public License v3.