Skip to content

testing.md

Latest commit

 

History

History
82 lines (61 loc) · 3.53 KB

testing.md

File metadata and controls

82 lines (61 loc) · 3.53 KB
Raw

Testing

Memory-limited runs with heavy-run

Activation adds scripts/ to PATH, which provides the heavy-run wrapper for workstation builds and tests with substantial memory use:

heavy-run ctest -j 1 --output-on-failure
HEAVY_RUN_MEM_HIGH=8G HEAVY_RUN_MEM_MAX=10G HEAVY_RUN_CPU=200% heavy-run make -j 2

The wrapper serialises jobs with the lock file /run/user/$UID/heavy-tests.lock (override with HEAVY_RUN_LOCK) and runs each job in a user systemd scope with MemoryHigh=12G, MemoryMax=16G and MemorySwapMax=2G. A CPU quota is optional. OpenMP and OpenBLAS thread counts default to 1 when they are not already set. heavy-run --help lists all overrides and the fallback lock path.

Behaviour without the required system support:

  • Without user systemd, heavy-run warns and serialises with flock only.
  • Without flock, it warns and runs without serialisation.
  • Memory and CPU limits require user systemd with delegated controllers.

Before starting the payload, the wrapper checks that it runs in the expected cgroup, that the memory controller is available and that the kernel's memory.max matches the configured MemoryMax. If a check fails, the payload is not run and the wrapper exits with status 125. HEAVY_RUN_MEM_MAX must be finite.

Tests for the wrapper itself:

bash tests/test_heavy_run.sh
bash tests/test_heavy_run.sh --memory   # Linux with user systemd only

The --memory variant allocates 1 GiB in Python under MemoryMax=512M, touching every page so that the allocation cannot be satisfied by the kernel zero page. It reports RSS, memory.current, swap usage and memory.events, and repeats the run with swap disabled; that run must end in an OOM kill with a max event. The test scopes use OOMPolicy=continue so that the final event counters can still be read.

Python integration tests

pytest tests/

The tests require the separate libneo Python package. The native _efit_to_boozer extension is optional; its tests are skipped when it is not available. Use --code-path /path/to/workspace (or set $CODE) to point to the workspace containing libneo; the default is the parent directory of this checkout.

Reference data

Several tests use private reference equilibria from gitlab.tugraz.at/plasma/data. Fetch selected subtrees with a read-only token:

export GITLAB_ACCESS_TOKEN=<read-only-token>
export DATA=$PWD/.testdata
scripts/fetch_data.sh AUG/EQDSK AUG/BOOZER/30835 DEMO/EQDSK MASTU/EQDSK \
    LHD/VMEC/makegrid_alternative TESTS/libneo/eqdsk
pytest tests/

The data location is taken from --data-path, then $DATA, then the checkout's .testdata directory. Tests whose data are missing are skipped.

Golden records are read-only by default, and tests with missing records are skipped. Use --regenerate-golden only after the new reference results have been checked independently.

Notes on scripts/fetch_data.sh:

  • Without a token it exits successfully without fetching anything.
  • With a token, clone, checkout and LFS errors are fatal; install git-lfs for LFS-tracked data.
  • In GitHub Actions it uses the data branch named like the current branch (GITHUB_HEAD_REF/GITHUB_REF_NAME), otherwise main, and falls back to main if that branch does not exist; DATA_BRANCH selects a branch explicitly.
  • Only the requested subtrees are copied; repeated fetches update them in place.

The Python integration suite and the data fetch are run manually. The GitHub workflows of this repository mirror it to GitLab and run the hygiene check; they do not run these tests.