Metrics & Logs

Metrics

Metrics queries GPU and host metrics. The schema is normalised, so the same metric names work for every provider in your fleet.

Building a query

1. Choose a Metric.
2. Choose an Agg (aggregation), for example avg.
3. Optionally select + filter to narrow the series.
4. Choose how to Group by: gpu, host, cluster or model.
5. Select Run query.

The query box shows the expression the builder produces, for example avg by (gpu) (gpu_utilization). Data is bucketed automatically for the chosen time range, for example into 15-second buckets over the last five minutes.

Available metrics

Metric	Measures
gpu_utilization	GPU utilisation.
gpu_memory_used	VRAM in use.
gpu_memory_pct	VRAM in use as a percentage of capacity.
gpu_temperature	GPU temperature.
gpu_power	GPU power draw.
sm_occupancy	Streaming multiprocessor occupancy.
host_cpu	Host CPU utilisation.
host_memory_used	Host memory in use.
host_load	Host load average.
host_net_rx, host_net_tx	Host network traffic received and transmitted.

Tip

gpu_utilization only shows whether a kernel was running, not how much of the GPU it used. Pair it with sm_occupancy to tell a truly busy GPU from one that is waiting on data loading.

Logs

Logs collects container, systemd and agent streams from across the fleet.

• Stream shows every line. Failures narrows the view to failures.
• Filter by severity: DEBUG, INFO, WARN, ERROR or CRIT. A counter above the list shows how many lines of each level match.
• Sort by Newest first, or change the order.
• Export downloads the matching lines.

If nothing matches, adjust the source, severity, search or time range. To see the logs of a single machine, use the Logs shortcut on its host page.