Skip to content

Machine usage report#

The machine usage report answers what did these machines do, and for whom? over a period you choose: one machine, a selection, a pool, or all of your organisation's machines on a cluster — including machines you lend to others, and your use of machines lent to you.

What is measured#

Every minute Astraeus adds what each machine did to that hour's record, kept 400 days. Each record is split by who: the workspace, the person, and the kind of work; what no one held is the machine's idle time.

Measure From
GPU-hours reserved GPUs held by work, times how long — whatever the work did with them.
GPU-hours used The GPUs' utilisation over the same time (a GPU busy half the time is half an hour per hour). GPU used is used over reserved.
GPU memory GiB in use × hours, and the peak.
CPU core-hours Reserved (cores promised to the work) and used (the work's CPU time).
Memory Reserved and used GiB × hours, and the peak.
Disk A machine's work and image filesystems; per workspace, the copies of its drives kept on the machine.
Network Bytes the work sent and received.
Energy kWh: the GPUs' own power readings (NVIDIA's, AMD's), integrated over time, plus an estimate for the CPU — 1 W per core at rest and 3 W more per busy core. Idle GPUs' draw counts in the machine's energy, not in anyone's.
Idle GPU-hours GPU-hours no one reserved.

Kinds of work: runs (a schedule's runs count as runs), deployments, agent runs, functions, notebooks and environments, other work (a replica group of yours), and drives (disk).

People: each workload is attributed to the person who made it — a run to whoever submitted it; a deployment's replicas, a function's, a schedule's runs, an agent's runs, a replica group's members to whoever made the deployment, function, schedule, agent or replica group. Work the cluster makes for itself (machine checks) is attributed to no one.

Who sees what#

You are You see
An organisation admin (or anyone with machines:read-metrics) Your machines whole: every workspace's and person's use — borrowers' included — and the idle time. On machines lent to you, your own use.
A workspace member Your workspace's use, wherever it ran. Your own work by name; other people's summed as Others, unless you may read machines' metrics (a workspace admin).
A lender Per lending: the borrowers' use by workspace and person. A person of another organisation shows as a person of (by name when they borrowed in person).
A borrower Per lending: your own use.

Open it#

Astraeus → Machines → Usage report, or the Usage section of a machine's page, or a lending's page (Organisation → Lendings).

  1. Choose the period: Today, Last 7 days, This month, Last month, or Custom dates.
  2. Narrow it to a machine, a pool, a person or a kind of work.
  3. Read the totals, the GPU-hours and energy by hour (two days or less) or by day, and the table by machine, workspace, person or kind of work. By person lists each person's kinds of work and largest workloads.
  4. Click CSV for every hour (or day), machine, workspace, lending, person and kind of work, with every measure.
$ astra astraeus usage-report --period last-month --by person
$ astra astraeus usage-report --machines gpu-1,gpu-2 --from 2026-10-01T00:00:00Z --csv > usage.csv
$ astra astraeus usage-report --pool h100 --kind deployment --by workspace

--period takes today, 7d, 30d, this-month, last-month; --by takes machine, workspace, person.

The organisation's machines (machines:read-metrics):

$ curl -sS "https://api.astralyx.cloud/v1/organizations/<org>/clusters/<cluster>/machine-usage?from=2026-10-01T00:00:00Z&to=2026-11-01T00:00:00Z&step=day" \
    -H "Authorization: Bearer $ASTRALYX_TOKEN" | jq '.totals'

Your workspace's own use: GET /v1/machine-usage with the same parameters. Filters: nodes (comma-separated), selector (pool=h100), namespace, person (a user id), kind, lending; format=csv for a CSV. Periods span at most 400 days (400 INVALID_PERIOD).

An AI assistant can read it too (get_machine_usage; get_lending_usage for a lending).

Notes#

  • A machine that stopped reporting (down, disconnected) is not measured for that time; what was reserved there still counts.
  • A gap in measurement longer than ten minutes (an outage) is left out: what happened then is not known.
  • Usage and cost is apart: what each workspace held, priced.