metric_sweep — one run, many result rows

Sources: examples/workflow_metric_sweep — settings in workflows/metric_sweep/. This one is not listed in the shipped workflow_settings.yml — add it to the workflows: list to run it:

workflows:
  - metric_sweep + channel_gain/*

Some tools do not produce a result, they produce a curve. A BER simulation sweeps its own Eb/N0 axis internally and writes one CSV row per point:

tasks:
  - name: main
    commands:
      - python3 simulate_ber.py --channel_gain ${channel_gain} --ebno_from 1 --ebno_to 3 --ebno_step 0.5

variables:
  channel_gain:
    type: list
    settings:
      list: [0.5, 1.0]

Two configurations, five Eb/N0 points each. Taking the first row of the CSV and discarding the rest would throw away most of the run, so multiple: true extracts every row, and a metadata block tags each one with the value it belongs to:

metrics:
  FER:
    type: csv
    multiple: true
    settings:
      file: results.csv
      key: FER

  BER:
    type: csv
    multiple: true
    settings:
      file: results.csv
      key: BER

  # "operation" metrics are evaluated once per expanded row, using that row's
  # own FER/BER values.
  fer_over_ber:
    type: operation
    settings:
      op: "FER / BER"

metadata:
  EBNO:
    type: csv
    multiple: true
    settings:
      file: results.csv
      key: EBNO

The two runs expand into ten result records, each carrying its own EBNO — as if each point had been run as a separate configuration, without paying the cost of ten separate jobs.

Two mechanisms worth separating:

• multiple: true turns one job into several records. metadata is what makes them distinguishable; without it, ten records would share the same identity.
• type: operation computes a metric from other metrics of the same record. Here it is evaluated per row, with that row’s own values — not once for the job.

Where to go next