Skip to content

Reading results

A run produces a self-contained, timestamped directory of provenance and data. This page covers what's in it, the CSV format, and the dashboard for visualizing results.

Output directory layout

Results are written under data/<CRAB_SYSTEM>/<name>_<timestamp>/, where <CRAB_SYSTEM> comes from the preset and <name> from global_options.name:

data/leonardo/interference_2026-06-09_14-30-05-123456/
├── config.json              # the exact config that ran (reproducibility)
├── environment.json         # the resolved environment
├── crab_job.sh              # the generated Slurm batch script
├── slurm_output.log         # job stdout
├── slurm_error.log          # job stderr
└── <experiment_name>/
    ├── data_app_0.csv        # collected metrics, one file per collected app
    ├── data_app_1.csv
    ├── error_app_<id>.log    # written only if an app failed
    └── run_<n>/.wrappers/    # per-run launch scripts (if retained)

A system-level data/<CRAB_SYSTEM>/metadata.csv registry is also appended to — one row per experiment (job name, timestamp, node count, ppn, app list, status, tags, relative path) — so you can index every run on a system from a single file. It's written with file locking, safe under concurrent CRAB jobs. See Architecture → Output layout.

config.json makes a run reproducible

Because the resolved config and environment are saved alongside the data, an output directory is a complete record of what ran. Keep it (or just config.json) to re-run the same study.

The CSV format

Each collected application produces data_app_<id>.csv. Columns:

  • run_id — which repetition the sample came from.
  • msg_size — message size parsed from the app's args (0 if not applicable).
  • one column per metric, named <app_id>_<metric_name>_<unit> (from the wrapper's metadata).

A wrapper that declares metrics performance (GTEPS) and time (s) for app 0 yields columns like 0_performance_GTEPS and 0_time_s. Multiple samples per run are stacked as separate rows sharing a run_id.

Load it straight into pandas:

import pandas as pd
df = pd.read_csv("data/leonardo/interference_.../a2a_vs_g500/data_app_0.csv")
df.groupby("run_id")["0_performance_GTEPS"].mean()

(Set outformat: "hdf" in global_options to get .h5 files instead of CSV.)

The dashboard

CRAB ships with a single-file HTML viewer bundled inside the package (src/crab/crab_dashboard.html) — no build step, just open it in a browser. Once loaded, it groups results by system and experiment, lets you toggle which columns and series are shown, choose axes (including violin plots), recolor series for comparison, and switch between dark and light themes. The Compare tab lets you select any combination of experiments — including across different labs/directories — and overlay or side-by-side them on a single chart, useful for eyeballing victim-vs-aggressor differences before deeper analysis in pandas or R.

You can get data into the dashboard three ways. For sharing results, use crab export — it bakes the data into the HTML itself.

The crab export command turns a results directory into a self-contained HTML file. The CSV data is embedded directly into the page, so the result is portable: email it, commit it, drop it on a web server — anyone with a browser can open it.

crab export data/leonardo/interference_2026-06-09_14-30-05-123456/ -o results.html

Default output is crab_export.html if you omit -o. The generated file should not be committed to version control — add it to your .gitignore. See CLI reference → crab export for full options.

Browse live results on the cluster

For exploring fresh results without exporting them, point the dashboard at an HTTP server running on the cluster. Start a simple server in the directory above your results and forward the port over SSH:

# on the cluster, in the parent of your data/ tree
python3 -m http.server 8000

# on your laptop
ssh -L 8000:localhost:8000 user@cluster

Then open src/crab/crab_dashboard.html from the cloned repo, enter the forwarded URL (e.g. http://localhost:8000/), and click Go. The dashboard crawls the directory listing for data_app_*.csv files across systems and experiments.

Drop a local folder

Drag a downloaded experiment folder onto the dashboard's drop zone (or click to browse). It reads the .csv files directly in your browser — nothing is uploaded anywhere.