3.3. Testing CICE

This section documents primarily how to use the CICE scripts to carry out CICE testing. Exactly what to test is a separate question and depends on the kinds of code changes being made. Prior to merging changes to the CICE Consortium main, changes will be reviewed and developers will need to provide a summary of the tests carried out.

There is a base suite of tests provided by default with CICE and this may be a good starting point for testing.

The testing scripts support several features

• Ability to test individual (via --test) or multiple tests (via --suite) using an input file to define the suite

• Ability to use test suites defined in the package or test suites defined by the user

• Ability to store test results for regresssion testing (--bgen)

• Ability to compare results to prior baselines to verify bit-for-bit (--bcmp)

• Ability to define where baseline tests are stored (--bdir)

• Ability to compare tests against each other (--diff)

• Ability to set or overide the batch account number (--acct) and queue name (--queue)

• Ability to control how test suites execute (--setup-only, --setup-build, --setup-build-run, --setup-build-submit)

3.3.1. Individual Tests

The CICE scripts support both setup of individual tests as well as test suites. Individual tests are run from the command line:

./cice.setup --test smoke --mach conrad --env cray --set diag1,debug --testid myid

Tests are just like cases but have some additional scripting around them. Individual tests can be created and manually modified just like cases. Many of the command line arguments for individual tests are similar to cice.setup Command Line Options for --case. For individual tests, the following command line options can be set

--test TESTNAME

specifies the test type. This is probably either smoke or restart but see cice.setup –help for the latest. This is required instead of --case.

--testid ID

specifies the testid. This is required for every use of --test and --suite. This is a user defined string that will allow each test to have a unique case and run directory name. This is also required.

--tdir PATH

specifies the test directory. Testcases will be created in this directory. (default is .)

--mach MACHINE (see cice.setup Command Line Options)

--env ENVIRONMENT1 (see cice.setup Command Line Options)

--set SET1,SET2,SET3 (see cice.setup Command Line Options)

--ignore-user-set (see cice.setup Command Line Options)

--acct ACCOUNT (see cice.setup Command Line Options)

--grid GRID (see cice.setup Command Line Options)

--pes MxNxBXxBYxMB (see cice.setup Command Line Options)

There are several additional options that come with --test that are not available with --case for regression and comparision testing,

--bdir DIR

specifies the top level location of the baseline results. This is used in conjuction with --bgen and --bcmp. The default is set by ICE_MACHINE_BASELINE in the env.[machine]_[environment] file.

--bgen DIR

specifies the name of the directory under [bdir] where test results will be stored. When this flag is set, it automatically creates that directory and stores results from the test under that directory. If DIR is set to default, then the scripts will automatically generate a directory name based on the CICE hash and the date and time. This can be useful for tracking the baselines by hash.

--bcmp DIR

specifies the name of the directory under [bdir] that the current tests will be compared to. When this flag is set, it automatically invokes regression testing and compares results from the current test to those prior results. If DIR is set to default, then the script will automatically generate the last directory name in the [bdir] directory. This can be useful for automated regression testing.

--diff LONG_TESTNAME

invokes a comparison against another local test. This allows different tests to be compared to each other for bit-for-bit-ness. This is different than --bcmp. --bcmp is regression testing, comparing identical test results between different model versions. --diff allows comparison of two different test cases against each other. For instance, different block sizes, decompositions, and other model features are expected to produced identical results and --diff supports that testing. The restrictions for use of --diff are that the test has to already be completed and the testid has to match. The LONG_TESTNAME string should be of format [test]_[grid]_[pes]_[sets]. The [machine], [env], and [testid] will be added to that string to complete the testname being compared. (See also Individual Test Examples #5)

The format of the case directory name for a test will always be [machine]_[env]_[test]_[grid]_[pes]_[sets].[testid] The [sets] will always be sorted alphabetically by the script so --set debug,diag1 and --set diag1,debug produces the same testname and test with _debug_diag1 in that order.

To build and run a test after invoking the ./cice.setup command, the process is the same as for a case. cd to the test directory, run the build script, and run the submit script:

cd [test_case]
./cice.build
./cice.submit

The test results will be generated in a local file called test_output. To check those results:

cat test_output

Tests are defined under configuration/scripts/tests/. Some tests currently supported are:

• smoke - Runs the model for default length. The length and options can

  be set with the --set command line option. The test passes if the model completes successfully.

• restart - Runs the model for 10 days, writing a restart file at the end of day 5 and

  again at the end of the run. Runs the model a second time starting from the day 5 restart and writes a restart at then end of day 10 of the model run. The test passes if both runs complete and if the restart files at the end of day 10 from both runs are bit-for-bit identical.

• decomp - Runs a set of different decompositions on a given configuration

Please run ./cice.setup --help for the latest information.

3.3.1.1. Adding a new test

See Test scripts

3.3.1.2. Individual Test Examples

1. Basic default single test

   Define the test, mach, env, and testid.

   ./cice.setup --test smoke --mach wolf --env gnu --testid t00
   cd wolf_gnu_smoke_col_1x1.t00
   ./cice.build
   ./cice.submit
   ./cat test_output
   

2. Simple test with some options

   Add --set

   ./cice.setup --test smoke --mach wolf --env gnu --set diag1,debug --testid t00
   cd wolf_gnu_smoke_col_1x1_debug_diag1.t00
   ./cice.build
   ./cice.submit
   ./cat test_output
   

3. Single test, generate a baseline dataset

   Add --bgen

   ./cice.setup --test smoke --mach wolf -env gnu --bgen cice.v01 --testid t00 --set diag1
   cd wolf_gnu_smoke_col_1x1_diag1.t00
   ./cice.build
   ./cice.submit
   ./cat test_output
   

4. Single test, compare results to a prior baseline

   Add --bcmp. For this to work, the prior baseline must exist and have the exact same base testname [machine]_[env]_[test]_[grid]_[pes]_[sets]

   ./cice.setup --test smoke --mach wolf -env gnu --bcmp cice.v01 --testid t01 --set diag1
   cd wolf_gnu_smoke_col_1x1_diag1.t01
   ./cice.build
   ./cice.submit
   ./cat test_output
   

5. Simple test, generate a baseline dataset and compare to a prior baseline

   Use --bgen and --bcmp. The prior baseline must exist already.

   ./cice.setup --test smoke --mach wolf -env gnu --bgen cice.v02 --bcmp cice.v01 --testid t02 --set diag1
   cd wolf_gnu_smoke_col_1x1_diag1.t02
   ./cice.build
   ./cice.submit
   ./cat test_output
   

6. Simple test, comparison against another test

   --diff provides a way to compare tests with each other. For this to work, the tests have to be run in a specific order and the testids need to match. The test is always compared relative to the current case directory.

   To run the first test,

   ./cice.setup --test smoke --mach wolf -env gnu --testid tx01 --set debug
   cd wolf_gnu_smoke_col_1x1_debug.tx01
   ./cice.build
   ./cice.submit
   ./cat test_output
   

   Then to run the second test and compare to the results from the first test

   ./cice.setup --test smoke --mach wolf -env gnu --testid tx01 --diff smoke_col_1x1_debug
   cd wolf_gnu_smoke_col_1x1.tx01
   ./cice.build
   ./cice.submit
   ./cat test_output
   

   The scripts will add a [machine]_[environment] to the beginning of the diff argument and the same testid to the end of the diff argument. Then the runs will be compared for bit-for-bit and a result will be produced in test_output.

3.3.1.3. Specific Test Cases

In addition to the test implemented in the general testing framework, specific tests have been developed to validate specific portions of the model. These specific tests are detailed in this section.

3.3.1.3.1. box2001

The box2001 test case is configured to perform the rectangular-grid box test detailed in [20]. It is configured to run a 72-hour simulation with thermodynamics disabled in a rectangular domain (80 x 80 grid cells) with a land boundary around the entire domain. It includes the following namelist modifications:

• dxrect: 16.e5 cm

• dyrect: 16.e5 cm

• ktherm: -1 (disables thermodynamics)

• coriolis: constant (f=1.46e-4 s\(^{-1}\))

• ice_data_type : box2001 (special initial ice mask)

• ice_data_conc : p5

• ice_data_dist : box2001 (special ice concentration initialization)

• atm_data_type : box2001 (special atmospheric and ocean forcing)

Ocean stresses are computed as in [20] where they are circular and centered in the square domain. The ice distribution is fixed, with a constant 2 meter ice thickness and a concentration field that varies linearly in the x-direction from 0 to 1 and is constant in the y-direction. No islands are included in this configuration. The test is configured to run on a single processor.

To run the test:

./cice.setup -m <machine> --test smoke -s box2001 --testid <test_id> --grid gbox80 --acct <queue manager account> -p 1x1

3.3.1.3.2. boxslotcyl

The boxslotcyl test case is an advection test configured to perform the slotted cylinder test detailed in [67]. It is configured to run a 12-day simulation with thermodynamics, ridging and dynamics disabled, in a square domain (80 x 80 grid cells) with a land boundary around the entire domain. It includes the following namelist modifications:

• dxrect: 10.e5 cm (10 km)

• dyrect: 10.e5 cm (10 km)

• ktherm: -1 (disables thermodynamics)

• kridge: -1 (disables ridging)

• kdyn: -1 (disables dynamics)

• ice_data_type : boxslotcyl (special initial ice mask)

• ice_data_conc : c1

• ice_data_dist : uniform

Dynamics is disabled because we directly impose a constant ice velocity. The ice velocity field is circular and centered in the square domain, and such that the slotted cylinder makes a complete revolution with a period \(T=\) 12 days :

(1)\[(u,v) = {u_0}\left( \frac{2y - L}{L}, \frac{-2x + L}{L}\right)\]

where \(L\) is the physical domain length and \(u_0 = \pi L / T\). The initial ice distribution is a slotted cylinder of radius \(r = 3L/10\) centered at \((x,y) = (L/2, 3L/4)\). The slot has a width of \(L/6\) and a depth of \(5L/6\) and is placed radially.

The time step is one hour, which with the above speed and mesh size yields a Courant number of 0.86.

The test can run on multiple processors.

To run the test:

./cice.setup -m <machine> --test smoke -s boxslotcyl --testid <test_id> --grid gbox80 --acct <queue manager account> -p nxm

3.3.2. Test suites

Test suites support running multiple tests specified via an input file. When invoking the test suite option (--suite) with cice.setup, all tests will be created, built, and submitted automatically under a local directory called testsuite.[testid] as part of involing the suite.:

./cice.setup --suite base_suite --mach wolf --env gnu --testid myid

Like an individual test, the --testid option must be specified and can be any string. Once the tests are complete, results can be checked by running the results.csh script in the testsuite.[testid]:

cd testsuite.[testid]
./results.csh

Multiple suites are supported on the command line as comma separated arguments:

./cice.setup --suite base_suite,decomp_suite --mach wolf --env gnu --testid myid

If a user adds --set to the suite, all tests in that suite will add that option:

./cice.setup --suite base_suite,decomp_suite --mach wolf --env gnu --testid myid -s debug

The option settings defined at the command line have precedence over the test suite values if there are conflicts.

The predefined test suites are defined under configuration/scripts/tests and the files defining the suites have a suffix of .ts in that directory. Some of the available tests suites are

quick_suite

consists of a handful of basic CICE tests

base_suite

consists of a much large suite of tests covering much of the CICE functionality

decomp_suite

checks that different decompositions and pe counts produce bit-for-bit results

omp_suite

checks that OpenMP single thread and multi-thread cases are bit-for-bit identical

io_suite

tests the various IO options including binary, netcdf, and pio. PIO should be installed locally and accessible to the CICE build system to make full use of this suite.

perf_suite

runs a series of tests to evaluate model scaling and performance

reprosum_suite

verifies that CICE log files are bit-for-bit with different decompositions and pe counts when the bfbflag is set to reprosum

gridsys_suite

tests B, C, and CD grid_ice configurations

prod_suite

consists of a handful of tests running 5 to 10 model years and includes some QC testing. These tests will be relatively expensive and take more time compared to other suites.

unittest_suite

runs unit tests in the CICE repository

travis_suite

consists of a small suite of tests suitable for running on low pe counts. This is the suite used with Github Actions for CI in the workflow.

first_suite

this small suite of tests is redundant with tests in other suites. It runs several of the critical baseline tests that other test compare to. It can improve testing turnaround if listed first in a series of test suites.

When running multiple suites on the command line (i.e. --suite first_suite,base_suite,omp_suite) the suites will be run in the order defined by the user and redundant tests across multiple suites will be created and executed only once.

The format for the test suite file is relatively simple. It is a text file with white space delimited columns that define a handful of values in a specific order. The first column is the test name, the second the grid, the third the pe count, the fourth column is the --set options and the fifth column is the --diff argument. The fourth and fifth columns are optional. Lines that begin with # or are blank are ignored. For example,

#Test   Grid  PEs  Sets                Diff
 smoke   col  1x1  diag1
 smoke   col  1x1  diag1,run1year  smoke_col_1x1_diag1
 smoke   col  1x1  debug,run1year
restart  col  1x1  debug
restart  col  1x1  diag1
restart  col  1x1  pondlvl
restart  col  1x1  pondtopo

The argument to --suite defines the test suite (.ts) filename and that argument can contain a path. cice.setup will look for the filename in the local directory, in configuration/scripts/tests/, or in the path defined by the --suite option.

Because many of the command line options are specified in the input file, ONLY the following options are valid for suites,

--suite filename

required, input filename with list of suites

--mach MACHINE

required

--env ENVIRONMENT1,ENVIRONMENT2

strongly recommended

--set SET1,SET2

optional

--acct ACCOUNT

optional

--tdir PATH

optional

--testid ID

required

--bdir DIR

optional, top level baselines directory and defined by default by ICE_MACHINE_BASELINE in env.[machine]_[environment].

--bgen DIR

recommended, test output is copied to this directory under [bdir]

--bcmp DIR

recommended, test output are compared to prior results in this directory under [bdir]

--report

This is only used by --suite and when set, invokes a script that sends the test results to the results page when all tests are complete. Please see Test Reporting for more information.

--coverage

When invoked, code coverage diagnostics are generated. This will modify the build and reduce optimization and generate coverage reports using lcov or codecov tools. General use is not recommended, this is mainly used as a diagnostic to periodically assess test coverage. Please see Code Coverage Testing for more information.

--setup-only

This is only used by --suite and when set, just creates the suite testcases. It does not build or submit them to run. By default, the suites do --setup-build-submit.

--setup-build

This is only used by --suite and when set, just creates and builds the suite testcases. It does not submit them to run. By default, the suites do --setup-build-submit.

--setup-build-run

This is only used by --suite and when set, runs the test cases interactively instead of submitting them in batch. By default, the suites do --setup-build-submit.

--setup-build-submit

This is only used by --suite and when set, sets up the cases, builds them, and submits them. This is the default behavior of suites.

Please see cice.setup Command Line Options and Individual Tests for more details about how these options are used.

As indicated above, cice.setup with --suite will create a directory called testsuite.[testid]. cice.setup also generates a script called suite.submit in that directory. suite.submit is the script that builds and submits the various test cases in the test suite.

The cice.setup* options --setup-only, --setup-build, and --setup-build-run modify how suite.submit is run by cice.setup. suite.submit can also be run manually, and the environment variables, SUITE_BUILD (builds the testcases), SUITE_RUN (runs the testcases interactively), and SUITE_SUBMIT (submit the testcases to run) control suite.submit. The default values for these variables are

SUITE_BUILD = true
SUITE_RUN = false
SUITE_SUBMIT = true

which means by default the test suite builds and submits the jobs. By defining other values for those environment variables, users can control the suite script. When using suite.submit manually, the string true (all lowercase) is the only string that will turn on a feature, and both SUITE_RUN and SUITE_SUBMIT cannot be true at the same time.

By leveraging the cice.setup command line arguments --setup-only, --setup-build, and --setup-build-run as well as the environment variables SUITE_BUILD, SUITE_RUN, and SUITE_SUBMIT, users can run cice.setup and suite.submit in various combinations to quickly setup, setup and build, submit, resubmit, run interactively, or rebuild and resubmit full testsuites quickly and easily. See Test Suite Examples for an example.

The script create_fails.csh will process the output from results.csh and generate a new test suite file, fails.ts, from the failed tests. fails.ts can then be edited and passed into cice.setup --suite fails.ts ... to rerun subsets of failed tests to more efficiently move thru the development, testing, and validation process. However, a full test suite should be run on the final development version of the code.

To report the test results, as is required for Pull Requests to be accepted into the main the CICE Consortium code see Test Reporting.

If using the --tdir option, that directory must not exist before the script is run. The tdir directory will be created by the script and it will be populated by all tests as well as scripts that support the test suite:

./cice.setup --suite base_suite --mach wolf --env gnu --testid myid --tdir /scratch/$user/testsuite.myid

3.3.2.1. Test Suite Examples

1. Basic test suite

   Specify suite, mach, env, testid.

   ./cice.setup --suite base_suite --mach conrad --env cray --testid v01a
   cd testsuite.v01a
   # wait for runs to complete
   ./results.csh
   

2. Basic test suite with user defined test directory

   Specify suite, mach, env, testid, tdir.

   ./cice.setup --suite base_suite --mach conrad --env cray --testid v01a --tdir /scratch/$user/ts.v01a
   cd /scratch/$user/ts.v01a
   # wait for runs to complete
   ./results.csh
   

3. Basic test suite on multiple environments

   Specify multiple envs.

   ./cice.setup --suite base_suite --mach conrad --env cray,pgi,intel,gnu --testid v01a
   cd testsuite.v01a
   # wait for runs to complete
   ./results.csh
   

   Each env can be run as a separate invokation of cice.setup but if that approach is taken, it is recommended that different testids be used.

4. Basic test suite with generate option defined

   Add --set

    ./cice.setup --suite base_suite --mach conrad --env gnu --testid v01b --set diag1
    cd testsuite.v01b
    # wait for runs to complete
   ./results.csh
   

   If there are conflicts between the --set options in the suite and on the command line, the command line options will take precedence.

5. Multiple test suites from a single command line

   Add comma delimited list of suites

   ./cice.setup --suite base_suite,decomp_suite --mach conrad --env gnu --testid v01c
   cd testsuite.v01c
   # wait for runs to complete
   ./results.csh
   

   If there are redundant tests in multiple suites, the scripts will understand that and only create one test.

6. Basic test suite, store baselines in user defined name

   Add --bgen

   ./cice.setup --suite base_suite --mach conrad --env cray --testid v01a --bgen cice.v01a
   cd testsuite.v01a
   # wait for runs to complete
   ./results.csh
   

   This will store the results in the default [bdir] directory under the subdirectory cice.v01a.

7. Basic test suite, store baselines in user defined top level directory

   Add --bgen and --bdir

   ./cice.setup --suite base_suite --mach conrad --env cray --testid v01a --bgen cice.v01a --bdir /tmp/user/CICE_BASELINES
   cd testsuite.v01a
   # wait for runs to complete
   ./results.csh
   

   This will store the results in /tmp/user/CICE_BASELINES/cice.v01a.

8. Basic test suite, store baselines in auto-generated directory

   Add --bgen default

   ./cice.setup --suite base_suite --mach conrad --env cray --testid v01a --bgen default
   cd testsuite.v01a
   # wait for runs to complete
   ./results.csh
   

   This will store the results in the default [bdir] directory under a directory name generated by the script that includes the hash and date.

9. Basic test suite, compare to prior baselines

   Add --bcmp

   ./cice.setup --suite base_suite --mach conrad --env cray --testid v02a --bcmp cice.v01a
   cd testsuite.v02a
   # wait for runs to complete
   ./results.csh
   

   This will compare to results saved in the baseline [bdir] directory under the subdirectory cice.v01a. With the --bcmp option, the results will be tested against prior baselines to verify bit-for-bit, which is an important step prior to approval of many (not all, see Code Validation Test (non bit-for-bit validation)) Pull Requests to incorporate code into the CICE Consortium main branch. You can use other regression options as well. (--bdir and --bgen)

10. Basic test suite, use of default string in regression testing

default is a special argument to --bgen and --bcmp. When used, the scripts will automate generation of the directories. In the case of --bgen, a unique directory name consisting of the hash and a date will be created. In the case of --bcmp, the latest directory in [bdir] will automatically be used. This provides a number of useful features

• the --bgen directory will be named after the hash automatically

• the --bcmp will always find the most recent set of baselines

• the --bcmp reporting will include information about the comparison directory name which will include hash information

• automation can be invoked easily, especially if --bdir is used to create separate baseline directories as needed.

Imagine the case where the default settings are used and --bdir is used to create a unique location. You could easily carry out regular builds automatically via,

set mydate = `date -u "+%Y%m%d"`
git clone https://github.com/myfork/cice cice.$mydate --recursive
cd cice.$mydate
./cice.setup --suite base_suite --mach conrad --env cray,gnu,intel,pgi --testid $mydate --bcmp default --bgen default --bdir /tmp/work/user/CICE_BASELINES_MAIN

When this is invoked, a new set of baselines will be generated and compared to the prior results each time without having to change the arguments.

11. Reusing a test suite

Add the buildincremental option (-s buildincremental). This permits the suite to be rerun without recompiling the whole code.

./cice.setup --suite base_suite --mach conrad --env intel --testid v01b --set buildincremental
cd testsuite.v01b
# wait for runs to complete
./results.csh
# modify code
./suite.submit
# wait for runs to complete
./results.csh

Only modified files will be recompiled, and the suite will be rerun.

12. Create and test a custom suite

Create your own input text file consisting of 5 columns of data,

• Test

• Grid

• pes

• sets (optional)

• diff test (optional)

such as

> cat mysuite
smoke    col  1x1  diag1,debug
restart  col  1x1
restart  col  1x1  diag1,debug    restart_col_1x1
restart  col  1x1  mynewoption,diag1,debug

then use that input file, mysuite

./cice.setup --suite mysuite --mach conrad --env cray --testid v01a --bgen default
cd testsuite.v01a
# wait for runs to complete
./results.csh

You can use all the standard regression testing options (--bgen, --bcmp, --bdir). Make sure any “diff” testing that goes on is on tests that are created earlier in the test list, as early as possible. Unfortunately, there is still no absolute guarantee the tests will be completed in the correct sequence.

13. Test suite generation then manual build followed by manual submission

Specify suite, mach, env, testid.

./cice.setup --suite quick_suite,base_suite --mach conrad --env cray,gnu --testid v01a --setup-only
cd testsuite.v01a
setenv SUITE_BUILD true
setenv SUITE_RUN false
setenv SUITE_SUBMIT false
./suite.submit
setenv SUITE_BUILD false
setenv SUITE_RUN false
setenv SUITE_SUBMIT true
./suite.submit
# wait for runs to complete
./results.csh

The setenv syntax is for csh/tcsh. In bash, the syntax would be SUITE_BUILD=true.

3.3.3. Unit Testing

Unit testing is supported in the CICE scripts. Unit tests are implemented via a distinct top level driver that tests CICE model features explicitly. These drivers can be found in cicecore/drivers/unittest/. In addition, there are some script files that also support the unit testing.

The unit tests build and run very much like the standard CICE model. A case is created and model output is saved to the case logs directory. Unit tests can be run as part of a test suite and the output is compared against an earlier set of output using a simple diff of the log files.

For example, to run the existing calendar unit test as a case,

./cice.setup -m onyx -e intel --case calchk01 -p 1x1 -s calchk
cd calchk01
./cice.build
./cice.submit

Or to run the existing calendar unit test as a test,

./cice.setup -m onyx -e intel --test unittest -p 1x1 --testid cc01 -s calchk --bgen cice.cc01
cd onyx_intel_unittest_gx3_1x1_calchk.cc01/
./cice.build
./cice.submit

To create a new unit test, add a new driver in cicecore/driver/unittest. The directory name should be the name of the test. Then create the appropriate set_nml or set_env files for the new unittest name in configuration/scripts/options. In particular, ICE_DRVOPT and ICE_TARGET need to be defined in a set_env file. Finally, edit configuration/scripts/Makefile and create a target for the unit test. The unit tests calchk or helloworld can be used as examples.

The following strings should be written to the log file at the end of the unit test run. The string “COMPLETED SUCCESSFULLY” will indicate the run ran to completion. The string “TEST COMPLETED SUCCESSFULLY” will indicate all the unit testing passed during the run. The unit test log file output is compared as part of regression testing. The string “RunningUnitTest” indicates the start of the output to compare. That string should be written to the log file at the start of the unit test model output. These strings will be queried by the testing scripts and will impact the test reporting. See other unit tests for examples about how these strings could be written.

The following are brief descriptions of some of the current unit tests,

• bcstchk is a unit test that exercises the methods in ice_broadcast.F90. This test does not depend on the CICE grid to carry out the testing. By testing with a serial and mpi configuration, both sets of software are tested independently and correctness is verified.

• calchk is a unit test that exercises the CICE calendar over 100,000 years and verifies correctness. This test does not depend on the CICE initialization.

• gridavgchk is a unit test that exercises the CICE grid_average_X2Y methods and verifies results.

• halochk is a unit test that exercises the CICE haloUpdate methods and verifies results.

• helloworld is a simple test that writes out helloworld and uses no CICE infrastructure. This tests exists to demonstrate how to build a unit test by specifying the object files directly in the Makefile

• optargs is a unit test that tests passing optional arguments down a calling tree and verifying that the optional attribute is preserved correctly.

• opticep is a cice test that turns off the icepack optional arguments passed into icepack. This can only be run with a subset of CICE/Icepack cases to verify the optional arguments are working correctly.

• sumchk is a unit test that exercises the methods in ice_global_reductions.F90. This test requires that a CICE grid and decomposition be initialized, so CICE_InitMod.F90 is leveraged to initialize the model prior to running a suite of unit validation tests to verify correctness.

3.3.4. Test Reporting

The CICE testing scripts have the capability to post test results to the official CICE Consortium Test-Results wiki page. You may need write permission on the wiki. If you are interested in using the wiki, please contact the Consortium. Note that in order for code to be accepted to the CICE main branch through a Pull Request it is necessary for the developer to provide proof that their code passes relevant tests. This can be accomplished by posting the full results to the wiki, or by copying the testing summary to the Pull Request comments.

To post results, once a test suite is complete, run results.csh and report_results.csh from the suite directory,

./cice.setup --suite base_suite --mach conrad --env cray --testid v01a
cd testsuite.v01a
#wait for runs to complete
./results.csh
./report_results.csh

report_results.csh will run results.csh by default automatically, but we recommmend running it manually first to verify results before publishing them. report_results.csh -n will turn off automatic running of results.csh.

The reporting can also be automated in a test suite by adding --report to cice.setup

./cice.setup --suite base_suite --mach conrad --env cray --testid v01a --report

With --report, the suite will create all the tests, build and submit them, wait for all runs to be complete, and run the results and report_results scripts.

3.3.5. Code Coverage Testing

The --coverage feature in cice.setup provides a method to diagnose code coverage. This argument turns on special compiler flags including reduced optimization and then invokes the gcov tool. Once runs are complete, either lcov or codecov can be used to analyze the results. This option is currently only available with the gnu compiler and on a few systems with modified Macros files. In the current implementation, when --coverage is invoked, the sandbox is copied to a new sandbox called something like cice_lcov_yymmdd-hhmmss. The source code in the new sandbox is modified slightly to improve coverage statistics and the full coverage suite is run there.

At the present time, the --coverage flag invokes the lcov analysis automatically by running the report_lcov.csh script in the test suite directory. The output will show up at the CICE lcov website. To use the tool, you should have write permission for that repository. The lcov tool should be run on a full multi-suite test suite, and it can take several hours to process the data once the test runs are complete. A typical instantiation would be

./cice.setup --suite first_suite,base_suite,travis_suite,decomp_suite,reprosum_suite,io_suite,quick_suite --mach cheyenne --env gnu --testid cc01 --coverage

Alternatively, codecov analysis can be carried out by manually running the report_codecov.csh script from the test suite directory, but there are several ongoing problems with this approach and it is not generally recommended. A script that summarizes the end-to-end process for codecov analysis can be found in ..**configuration/scripts/tests/cice_test_codecov.csh**. The codecov analysis is largely identical to the analysis performed by lcov, codecov just provides a nicer web experience to view the output.

This is a special diagnostic test and is not part of the standard model testing. General use is not recommended, this is mainly used as a diagnostic to periodically assess test coverage.

..Because codecov.io does not support git submodule analysis right now, a customized ..repository has to be created to test CICE with Icepack integrated directly. The repository ..https://github.com/apcraig/Test_CICE_Icepack serves as the current default test repository. ..In general, to setup the code coverage test in CICE, the current CICE main has ..to be copied into the Test_CICE_Icepack repository, then the full test suite ..can be run with the gnu compiler with the --coverage argument.

..The test suite will run and then a report will be generated and uploaded to ..the codecov.io site by the ..**report_codecov.csh** script. The env variable CODECOV_TOKEN needs to be defined ..either in the environment or in a file named ~/.codecov_cice_token. That ..token provides write permission to the Test_CICE_Icepack codecov.io site and is available ..by contacting the Consortium team directly.

..A script that carries out the end-to-end testing can be found in ..**configuration/scripts/tests/cice_test_codecov.csh**

..This is a special diagnostic test and does not constitute proper model testing. ..General use is not recommended, this is mainly used as a diagnostic to periodically ..assess test coverage. The interaction with codecov.io is not always robust and ..can be tricky to manage. Some constraints are that the output generated at runtime ..is copied into the directory where compilation took place. That means each ..test should be compiled separately. Tests that invoke multiple runs ..(such as exact restart and the decomp test) will only save coverage information ..for the last run, so some coverage information may be lost. The gcov tool can ..be a little slow to run on large test suites, and the codecov.io bash uploader ..(that runs gcov and uploads the data to codecov.io) is constantly evolving. ..Finally, gcov requires that the diagnostic output be copied into the git sandbox for ..analysis. These constraints are handled by the current scripts, but may change ..in the future.

3.3.6. Code Validation Test (non bit-for-bit validation)

A core tenet of CICE dycore and CICE innovations is that they must not change the physics and biogeochemistry of existing model configurations, notwithstanding obsolete model components. Therefore, alterations to existing CICE Consortium code must only fix demonstrable numerical or scientific inaccuracies or bugs, or be necessary to introduce new science into the code. New physics and biogeochemistry introduced into the model must not change model answers when switched off, and in that case CICEcore and CICE must reproduce answers bit-for-bit as compared to previous simulations with the same namelist configurations. This bit-for-bit requirement is common in Earth System Modeling projects, but often cannot be achieved in practice because model additions may require changes to existing code. In this circumstance, bit-for-bit reproducibility using one compiler may not be unachievable on a different computing platform with a different compiler. Therefore, tools for scientific testing of CICE code changes have been developed to accompany bit-for-bit testing. These tools exploit the statistical properties of simulated sea ice thickness to confirm or deny the null hypothesis, which is that new additions to the CICE dycore and CICE have not significantly altered simulated ice volume using previous model configurations. Here we describe the CICE testing tools, which are applies to output from five-year gx-1 simulations that use the standard CICE atmospheric forcing. A scientific justification of the testing is provided in [24]. The following sections follow [49].

3.3.6.1. Two-Stage Paired Thickness Test

The first quality check aims to confirm the null hypotheses \(H_0\!:\!\mu_d{=}0\) at every model grid point, given the mean thickness difference \(\mu_d\) between paired CICE simulations ‘\(a\)’ and ‘\(b\)’ that should be identical. \(\mu_d\) is approximated as \(\bar{h}_{d}=\tfrac{1}{n}\sum_{i=1}^n (h_{ai}{-}h_{bi})\) for \(n\) paired samples of ice thickness \(h_{ai}\) and \(h_{bi}\) in each grid cell of the gx-1 mesh. Following [66], the associated \(t\)-statistic expects a zero mean, and is therefore

(2)\[t=\frac{\bar{h}_{d}}{\sigma_d/\sqrt{n_{eff}}}\]

given variance \(\sigma_d^{\;2}=\frac{1}{n-1}\sum_{i=1}^{n}(h_{di}-\bar{h}_d)^2\) of \(h_{di}{=}(h_{ai}{-}h_{bi})\) and effective sample size

(3)\[n_{eff}{=}n\frac{({1-r_1})}{({1+r_1})}\]

for lag-1 autocorrelation:

(4)\[r_1=\frac{\sum\limits_{i=1}^{n-1}\big[(h_{di}-\bar{h}_{d1:n-1})(h_{di+1}-\bar{h}_{d2:n})\big]}{\sqrt{\sum\limits_{i=1}^{n-1} (h_{di}-\bar{h}_{d1:n-1})^2 \sum\limits_{i=2}^{n} (h_{di}-\bar{h}_{d2:n})^2 }}.\]

Here, \(\bar{h}_{d1:n-1}\) is the mean of all samples except the last, and \(\bar{h}_{d2:n}\) is the mean of samples except the first, and both differ from the overall mean \(\bar{h}_d\) in equations ((2)). That is:

(5)\[\bar{h}_{d1:n-1}=\frac{1}{n{-}1} \sum \limits_{i=1}^{n-1} h_{di},\quad \bar{h}_{d2:n}=\frac{1}{n{-}1} \sum \limits_{i=2}^{n} h_{di},\quad \bar{h}_d=\frac{1}{n} \sum \limits_{i=1}^{n} {h}_{di}\]

Following [68], the effective sample size is limited to \(n_{eff}\in[2,n]\). This definition of \(n_{eff}\) assumes ice thickness evolves as an AR(1) process [62], which can be justified by analyzing the spectral density of daily samples of ice thickness from 5-year records in CICE Consortium member models [24]. The AR(1) approximation is inadmissible for paired velocity samples, because ice drift possesses periodicity from inertia and tides [18][36][50]. Conversely, tests of paired ice concentration samples may be less sensitive to ice drift than ice thickness. In short, ice thickness is the best variable for CICE Consortium quality control (QC), and for the test of the mean in particular.

Care is required in analyzing mean sea ice thickness changes using ((2)) with \(N{=}n_{eff}{-}1\) degrees of freedom. [68] demonstrate that the \(t\)-test in ((2)) becomes conservative when \(n_{eff} < 30\), meaning that \(H_0\) may be erroneously confirmed for highly auto-correlated series. Strong autocorrelation frequently occurs in modeled sea ice thickness, and \(r_1>0.99\) is possible in parts of the gx-1 domain for the five-year QC simulations. In the event that \(H_0\) is confirmed but \(2\leq n_{eff}<30\), the \(t\)-test progresses to the ‘Table Lookup Test’ of [68], to check that the first-stage test using ((2)) was not conservative. The Table Lookup Test chooses critical \(t\) values \(|t|<t_{crit}({1{-}\alpha/2},N)\) at the \(\alpha\) significance level based on \(r_1\). It uses the conventional \(t={\bar{h}_{d} \sqrt{n}}/{\sigma_d}\) statistic with degrees of freedom \(N{=}n{-}1\), but with \(t_{crit}\) values generated using the Monte Carlo technique described in [68], and summarized in Two-sided t_{crit} values for 5-year QC simulations (\(N=1824\)) at the two-sided 80% confidence interval (\(\alpha=0.2\)). We choose this interval to limit Type II errors, whereby a QC test erroneously confirms \(H_0\).

Table Two-sided t_{crit} values shows the summary of two-sided \(t_{crit}\) values for the Table Lookup Test of [68] at the 80% confidence interval generated for \(N=1824\) degrees of freedom and lag-1 autocorrelation \(r_1\).

Two-sided \(t_{crit}\) values

\(r_1\)	-0.05	0.0	0.2	0.4	0.5	0.6	0.7	0.8	0.9	0.95	0.97	0.99
\(t_{crit}\)	1.32	1.32	1.54	2.02	2.29	2.46	3.17	3.99	5.59	8.44	10.85	20.44

3.3.6.2. Quadratic Skill Validation Test

In addition to the two-stage test of mean sea ice thickness, we also check that paired simulations are highly correlated and have similar variance using a skill metric adapted from [58]. A general skill score applicable to Taylor diagrams takes the form

(6)\[S_m=\frac{4(1+R)^m}{({\hat{\sigma}_{f}+1/{\hat{\sigma}_{f}}})^2 (1+R_0)^m}\]

where \(m=1\) for variance-weighted skill, and \(m=4\) for correlation-weighted performance, as given in equations (4) and (5) of [58], respectively. We choose \(m=2\) to balance the importance of variance and correlation reproduction in QC tests, where \(\hat{\sigma}_{f}={\sigma_{b}}/{\sigma_{a}}\) is the ratio of the standard deviations of simulations ‘\(b\)’ and ‘\(a\)’, respectively, and simulation ‘\(a\)’ is the control. \(R_0\) is the maximum possible correlation between two series for correlation coefficient \(R\) calculated between respective thickness pairs \(h_{a}\) and \(h_{b}\). Bit-for-bit reproduction of previous CICE simulations means that perfect correlation is possible, and so \(R_0=1\), giving the quadratic skill of run ‘\(b\)’ relative to run ‘\(a\)’:

(7)\[S=\bigg[ \frac{(1+R) (\sigma_a \sigma_b)}{({\sigma_a}^2 + {\sigma_b}^2)} \bigg]^2\]

This provides a skill score between 0 and 1. We apply this \(S\) metric separately to the northern and southern hemispheres of the gx-1 grid by area-weighting the daily thickness samples discussed in the Two-Stage Paired Thickness QC Test. The hemispheric mean thickness over a 5-year simulation for run ‘\(a\)’ is:

(8)\[\bar{h}_{a}=\frac{1}{n} \sum_{i=1}^{n} \sum_{j=1}^{J} \ W_{j} \; h_{{a}_{i,j}}\]

at time sample \(i\) and grid point index \(j\), with an equivalent equation for simulation ‘\(b\)’. \(n\) is the total number of time samples (nominally \(n=1825\)) and \(J\) is the total number of grid points on the gx-1 grid. \(W_j\) is the weight attributed to each grid point according to its area \(A_{j}\), given as

(9)\[W_{j}=\frac{ A_{j} }{\sum_{j=1}^{J} A_{j}}\]

for all grid points within each hemisphere with one or more non-zero thicknesses in one or both sets of samples \(h_{{a}_{i,j}}\) or \(h_{{b}_{i,j}}\). The area-weighted variance for simulation ‘\(a\)’ is:

(10)\[\sigma_a^{\;2}=\frac{\hat{J}}{(n\,\hat{J}-1)} \sum_{i=1}^{n} \sum_{j=1}^{J} W_{j} \, (h_{{a}_{i,j}}-\bar{h}_{a})^2\]

where \(\hat{J}\) is the number of non-zero \(W_j\) weights, and \(\sigma_b\) is calculated equivalently for run ‘\(b\)’. In this context, \(R\) becomes a weighted correlation coefficient, calculated as

(11)\[R=\frac{\textrm{cov}(h_{a},h_{b})}{\sigma_a \; \sigma_b}\]

given the weighted covariance

(12)\[\textrm{cov}(h_{a},h_{b})=\frac{\hat{J}}{(n\,\hat{J}-1)} \sum_{i=1}^{n} \sum_{j=1}^{J} W_{j} \, (h_{{a}_{i,j}}-\bar{h}_{a}) (h_{{b}_{i,j}}-\bar{h}_{b}).\]

Using equations ((7)) to ((12)), the skill score \(S\) is calculated separately for the northern and southern hemispheres, and must exceed a critical value nominally set to \(S_{crit}=0.99\) to pass the test. Practical illustrations of this test and the Two-Stage test described in the previous section are provided in [24].

3.3.6.3. Code Validation Testing Procedure

The CICE code validation (QC) test is performed by running a python script (configurations/scripts/tests/QC/cice.t-test.py). In order to run the script, the following requirements must be met:

• Python v2.7 or later

• netcdf Python package

• numpy Python package

• matplotlib Python package (optional)

• basemap Python package (optional)

QC testing should be carried out using configurations (ie. namelist settings) that exercise the active code modifications. Multiple configurations may need to be tested in some cases. Developers can contact the Consortium for guidance or if there are questions.

In order to generate the files necessary for the validation test, test cases should be created with the qc option (i.e., --set qc) when running cice.setup. This option results in daily, non-averaged history files being written for a 5 year simulation.

To install the necessary Python packages, the pip Python utility can be used.

pip install --user netCDF4
pip install --user numpy
pip install --user matplotlib
pip install --user cartopy

You can also setup a conda env with the same utitities

conda env create -f configuration/scripts/tests/qctest.yml
conda activate qctest

To run the validation test, setup a baseline run with the original baseline model and then a perturbation run based on recent model changes. Use --set qc in both runs in addition to other settings needed. Then use the QC script to compare history output,

cp configuration/scripts/tests/QC/cice.t-test.py .
./cice.t-test.py /path/to/baseline/history /path/to/test/history

The script will produce output similar to:

INFO:__main__:Number of files: 1825

INFO:__main__:Two-Stage Test Passed

INFO:__main__:Quadratic Skill Test Passed for Northern Hemisphere

INFO:__main__:Quadratic Skill Test Passed for Southern Hemisphere

INFO:__main__:

INFO:__main__:Quality Control Test PASSED

Additionally, the exit code from the test (echo $?) will be 0 if the test passed, and 1 if the test failed.

The cice.t-test.py requires memory to store multiple two-dimensional fields spanning 1825 unique timesteps, a total of several GB. An appropriate resource is needed to run the script. If the script runs out of memory on an interactive resource, try logging into a batch resource or finding a large memory node.

The cice.t-test.py script will also attempt to generate plots of the mean ice thickness for both the baseline and test cases. Additionally, if the 2-stage test fails then the script will attempt to plot a map showing the grid cells that failed the test. For a full list of options, run python cice.t-test.py -h.

3.3.6.4. End-To-End Testing Procedure

Below is an example of a step-by-step procedure for testing a code change that might result in non bit-for-bit results. First, run a regression test,

# Run a full regression test to verify bit-for-bit

# Create a baseline dataset (only necessary if no baseline exists on the system)
# if you want to replace an existing baseline, you should first delete the directory cice.my.baseline in ${ICE_BASELINE}.
# git clone the baseline code

./cice.setup -m onyx -e intel --suite base_suite --testid base0 --bgen cice.my.baseline

# Check the results

cd testsuite.base0
./results.csh

# Run the test suite with the new code
# git clone the new code

./cice.setup -m onyx -e intel --suite base_suite --testid test0 --bcmp cice.my.baseline

# Check the results

cd testsuite.test0
./results.csh

# Note which tests failed and determine which namelist options are responsible for the failures

If the regression comparisons fail, then you may want to run the QC test,

# Run the QC test

# Create a QC baseline
# From the baseline sandbox
# Generate the test case(s) using options or namelist changes to activate new code modifications

./cice.setup -m onyx -e intel --test smoke -g gx1 -p 44x1 --testid qc_base -s qc,medium
cd onyx_intel_smoke_gx1_44x1_medium_qc.qc_base
# modify ice_in to activate the namelist options that were determined above
./cice.build
./cice.submit

# Create the t-test testing data
# From the updated sandbox
# Generate the same test case(s) as the baseline using options or namelist changes to activate new code modifications

./cice.setup -m onyx -e intel --test smoke -g gx1 -p 44x1 --testid qc_test -s qc,medium
cd onyx_intel_smoke_gx1_44x1_medium_qc.qc_test
# modify ice_in to activate the namelist options that were determined above
./cice.build
./cice.submit

# Wait for runs to finish
# Perform the QC test

# From the updated sandbox
cp configuration/scripts/tests/QC/cice.t-test.py .
./cice.t-test.py /p/work/turner/CICE_RUNS/onyx_intel_smoke_gx1_44x1_medium_qc.qc_base \
                 /p/work/turner/CICE_RUNS/onyx_intel_smoke_gx1_44x1_medium_qc.qc_test

# Example output:
INFO:__main__:Number of files: 1825
INFO:__main__:Two-Stage Test Passed
INFO:__main__:Quadratic Skill Test Passed for Northern Hemisphere
INFO:__main__:Quadratic Skill Test Passed for Southern Hemisphere
INFO:__main__:
INFO:__main__:Quality Control Test PASSED