Skip to content

DOCKER.md

Latest commit

 

History

History
135 lines (106 loc) · 5.53 KB

DOCKER.md

File metadata and controls

135 lines (106 loc) · 5.53 KB
Raw

Running TFBoost in a container

A single image that bundles everything TFBoost needs — ROOT (from the base image, not recompiled), the HYDRA headers, TBB, TCLAP, libconfig and FFTW — builds the TBB backends, and runs a chosen target against input data you mount at /data.

Works with both podman and docker (the commands below use podman; swap in docker if you prefer).

Build the image

Run from the repository root (the build context must be the repo, while the Dockerfile lives in install/):

podman build -f install/Dockerfile -t tfboost .

This pulls a ROOT base image, clones HYDRA, builds: analysis_tbb deconvolution_tbb resampling_tbb 3Ddiamond_tbb analyze_single_signal_tbb simple_analysis_tbb.

Build options (--build-arg)

Arg Default Purpose
BASE_IMAGE rootproject/root:latest ROOT-carrying base image; pin a tag for reproducibility.
HYDRA_REPO …/MultithreadCorner/Hydra.git HYDRA source repo.
HYDRA_REF master HYDRA branch/tag (verified build used 4.0.1).
TFB_MARCH native CPU target. See portability note below.
TFB_TARGETS the 6 targets above Space-separated list of targets to compile.

Portability note. CMakeLists.txt forces -march=native, so the binaries are tuned to the build host's CPU and may crash with illegal instruction on a different machine. For a same-machine image the default is fine. To make a portable image, build with e.g. --build-arg TFB_MARCH=x86-64-v2 (the Dockerfile patches the flag inside the image only).

Run

Mount a host folder at /data. Outputs are written under /data/results by default, so they persist after the container exits.

# Default: runs analysis_tbb. Put input signals in ./data/input to have them
# picked up automatically; results appear in ./data/results.
podman run --rm -v "$PWD/data:/data" tfboost
# Point at an arbitrary input directory inside the mount:
podman run --rm -v "$PWD/data:/data" -e TFB_INPUT=/data/signals tfboost

# Run a different target:
podman run --rm -v "$PWD/data:/data" tfboost deconvolution_tbb

# Use your own config file (mounted in):
podman run --rm -v "$PWD/data:/data" -e TFB_CONFIG=/data/my.cfg tfboost

# Drop into a shell for debugging:
podman run --rm -it -v "$PWD/data:/data" tfboost shell

Launching the GUI from the container

The image also carries the Python/Tk front-end. ./start.sh is the easy path — it detects the runtime, builds the image and wires up X11 for you. To do it by hand, the container needs both the X socket and an authorization cookie. Passing the socket alone gives "Authorization required, but no authorization protocol specified", because the X server wants a cookie. The portable fix is to export an Xauthority entry with a wildcard hostname:

# Build a cookie the container can use (wildcard hostname => matches inside it):
XAUTH=$(mktemp)
xauth nlist "$DISPLAY" | sed -e 's/^..../ffff/' | xauth -f "$XAUTH" nmerge -
chmod 644 "$XAUTH"

podman run --rm \
    -e DISPLAY \
    -e XAUTHORITY=/tmp/.Xauthority \
    -v "$XAUTH:/tmp/.Xauthority:ro" \
    -v /tmp/.X11-unix:/tmp/.X11-unix \
    --security-opt label=disable \
    -v "$PWD/data:/data" \
    tfboost gui

rm -f "$XAUTH"

Notes / troubleshooting:

  • The sed 's/^..../ffff/' rewrites the cookie's address family to FamilyWild, so it authenticates regardless of the hostname the container presents — this is what fixes the "Authorization required" error.
  • --security-opt label=disable lets a rootless-podman container reach the X socket and the mounted cookie on SELinux systems (Fedora); on Docker/non-SELinux it is a harmless no-op.
  • On Wayland this uses XWayland (the /tmp/.X11-unix socket), which works for Tkinter. If windows still do not appear, check echo $DISPLAY is set on the host.
  • Window too small / cramped on a HiDPI or fractionally-scaled display? The GUI has a fixed-pixel layout designed for ~96 dpi, and XWayland reports a fake 96 dpi so nothing auto-scales. Pass -e TFB_UI_SCALE=1.5 (or 1.25, 2.0, …) to scale the whole UI — window, widgets and fonts — uniformly. start.sh sets this automatically from your largest configured monitor scale (~/.config/monitors.xml); override with TFB_UI_SCALE=1.25 ./start.sh.
  • ./start.sh does all of the above automatically (X11 auth + UI scaling, with xhost +local: as an extra fallback) — prefer it.

Runtime environment variables

Var Default Effect
TFB_TARGET analysis_tbb (or 1st CLI arg) Which executable to run.
TFB_CONFIG baked example config Config file to use; copied into the path the binary reads.
TFB_INPUT /data/input if non-empty, else the config value Input-signals directory.
TFB_OUTPUT /data/results Output directory (plots/ + data/ are created inside).

The entrypoint rewrites InputDirectory / OutputDirectory in the active configuration.cfg (the file analysis_tbb reads). Other targets read their own etc/*.cfg; for those, supply paths via TFB_CONFIG or mount an edited etc/.

With no /data mount and no overrides the image still runs end-to-end against the bundled examples/input_files, writing to /data/results inside the (ephemeral) container — handy as a smoke test:

podman run --rm tfboost

Verified toolchain

The reference build this image mirrors used ROOT 6.40.02, HYDRA 4.0.1 and the system GCC with --std=c++20. Any ROOT ≥ 6.14 in the base image is expected to work.