BUILD SERVER UTILIZATION PREDICTOR¶
Trains a daily-utilization regression model for one HPC server using the last N days of hpcjobs history (default 180). Persists the model artefact under <server_folder>/ml/utilization_predictor.json and returns a 14-day forecast (mean + 2σ bands), trend direction (up/down/stable), MAPE on a 14-day holdout, R², and the historical daily series. Drives the trend predictor on the Modern utilization tab.
When to use¶
Tagged: hpcserver, ml, regression, trend, forecast, utilization, platform.utilization.
Inputs¶
| Label | ID | Type | Default | Required | Description |
|---|---|---|---|---|---|
| HPC Server | hpcserver_id | remote_lookup | — | ✓ | The HPC server whose utilization history will be fit. The server’s admin folder is auto-created if missing; the model lands under <admin_dir>/ml/utilization_predictor.json. |
| Window (days) | window_days | number | 180 | How many days of daily-aggregated hpcjobs history to fit on. Minimum 14, maximum 730. A 180-day window balances recency against stability. | |
| Forecast horizon (days) | forecast_days | number | 14 | How many days into the future to predict. Each forecast point includes a mean and a ±2σ band derived from in-sample residuals. Maximum 90. |
Outputs¶
| Label | ID | Type | Description |
|---|---|---|---|
| Status | status | string | |
| Model Path | model_path | string | |
| Trend | trend_direction | string | |
| Slope (jobs/day) | trend_slope | scalar | |
| MAPE % | mape_percent | scalar | |
| R² | r_squared | scalar | |
| Mean Jobs/Day | mean_jobs_per_day | scalar | |
| History | history | dataset | |
| Forecast | forecast | dataset |
Auto-generated from platform schema. Worker id: hpcserver_ml_builder. Schema hash: 5c296f354c7a. Hand-curated docs in workerexamples/ override this page when present.