Add JSON output for benchmark results

[dantegd]

Builds on top of #7980

This PR adds a structured JSON output path for cuML benchmarks. Benchmark runs can now write a single canonical JSON artifact containing grouped benchmark results and run metadata, including backend timings, accuracy metrics, benchmark shape, selected config/profile information, hardware metadata, package environment snapshots, and estimator parameters when available.

This PR also improves benchmark console output for long runs. Instead of repeated per-run DataFrame dumps, the runner now emits a concise aligned progress table with algorithm, shape, estimated data size, backend timings, and compact details like accuracy and speedup.

Why JSON Output

I think YAML remains the right format for benchmark configuration because it is human-authored, reviewable, and easy to maintain. JSON is the right format for benchmark output because it is machine-produced, schema-versioned, and easy for CI, dashboards, and future baseline comparison tooling to consume.

YAML in: human-authored benchmark intent
JSON out: machine-readable benchmark facts

JSON Schema

The JSON artifact has two top-level fields:

{
  "results": [],
  "metadata": {}
}

Each result represents one logical benchmark variation:

{
  "benchmark_id": "logreg_fit_narrow_default",
  "algorithm": "LogisticRegression",
  "dataset": "classification",
  "operation": "fit",
  "shape": {
    "rows": 84000000,
    "features": 16,
    "estimated_input_size_bytes": 5376000000,
    "estimated_input_size_gb": 5.376
  },
  "data": {
    "input_type": "cupy",
    "dtype": "fp32",
    "n_reps": 1
  },
  "params": {
    "declared": {},
    "effective": {
      "cpu": null,
      "gpu": {},
      "accel": null
    }
  },
  "backends": {
    "gpu": {
      "status": "success",
      "time_sec": 0.91,
      "accuracy": 0.995
    }
  }
}

Run metadata includes schema version, command, Python/platform information, cuML/git identity, runtime availability, config/profile selection, hardware details, and a compact conda or pip package snapshot.

Backend records currently use:

• status: success or skipped
• time_sec: benchmark time in seconds for successful runs
• accuracy: accuracy or score when available
• reason: explanation for skipped backends

Highlights

• Adds --output results.json as the canonical benchmark artifact.
• Writes a single JSON file with top-level results followed by metadata.
• Groups backend results under each logical benchmark variation.
• Captures backend status for successful and skipped runs.
• Captures declared parameters and effective estimator parameters when available.
• Captures run metadata including command, config/profile, Python/platform, cuML/git info, GPU availability, hardware details, and compact conda/pip package snapshots.
• Adds hardware metadata overrides for cleaner reporting labels.
• Keeps --csv as a legacy flat export.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.