Environment Bundles — Overview¶

AOBench ships 26 deterministic HPC environment snapshot bundles under benchmark/environments/env_01/ … env_26/, plus six M100 ExaData-grounded bundles (env_m100_01 … env_m100_06, documented in their own section below). Each bundle freezes a realistic operational scenario — a job failure, a queue-congestion event, a cooling unit fault, a policy violation, a multi-job interference incident, and so on — so that any agent run against the bundle is reproducible.

This page is the cross-reference index: scenario type, scored roles, scored QCATs, and the human description for each bundle. For the snapshot file format (slurm_state.json, telemetry/*.parquet, rbac_policy.yaml, docs/*.md, incident_metadata.json), see environments.md. For the authoritative metadata.yaml of each bundle, read the file directly.

Scenario coverage at a glance¶

All 23 bundles¶

M100 ExaData-grounded bundles¶

A separate set of six bundles grounded in the real CINECA Marconi100 (M100) ExaData dataset. Telemetry uses real M100 metric vocabulary (r{rack}n{slot} nodes, IPMI/ganglia/ vertiv metric names) and values sampled from real M100 distributions — or real per-node traces in --real-baselines mode — with a controlled, labeled perturbation per scenario. All six pass the full F1–F7 fidelity gate.

Rebuild with scripts/build_m100_bundles.py (deterministic; --real-baselines for real per-node traces, --relative-anomalies to scale anomalies to each node's real load). A gold-consistency guard runs at the end of the build. See the M100 environments guide for details.

Per-bundle file layout¶

Every bundle has the same shape (some optional files appear only when relevant to the scenario):

benchmark/environments/env_NN/
├── metadata.yaml                       Bundle metadata (env_id, scenario_type,
│                                       supported_roles, supported_categories,
│                                       included_sources)
├── manifest.txt                        Sorted list of all included files
├── slurm/
│   ├── slurm_state.json                Nodes, partitions, jobs (validated by SlurmState)
│   └── job_details.json                sacct-level details (when relevant)
├── telemetry/
│   ├── telemetry_timeseries.parquet    columns: timestamp, node_id, metric_name, value, unit
│   └── memory_events.csv               OOM / memory events (when relevant)
├── policy/
│   └── rbac_policy.yaml                Per-role permissions (schema v1.1)
├── docs/
│   ├── *.md                            User-facing knowledge documents
│   └── rbac_policy.md                  Auto-generated readable policy summary
└── incidents/
    └── incident_metadata.json          Incident timeline + affected resources

The metadata.yaml of every bundle is the authoritative source for what is in scope. To list every file manifest.txt records:

cat benchmark/environments/env_05/manifest.txt

Validating a bundle¶

aobench validate benchmark runs validate_bundle() from src/aobench/environment/snapshot_validator.py over every bundle. It checks JSON-schema conformance for slurm_state.json, incident_metadata.json, and rbac_policy.yaml, plus the parquet column schema for telemetry/*.parquet.

To regenerate a bundle (for example after editing the source CSVs):

python scripts/generate_bundles.py --env env_NN

scripts/generate_bundles.py also emits the auto-generated docs/rbac_policy.md per environment from the YAML policy.