> Index of every page on this site: [llms.txt](https://docs.a-line.cloud/llms.txt)

# Runner sizing

> How A-Line.cloud measures each Actions job and suggests a bigger or smaller runner.

Every GitHub Actions job that runs on an A-Line.cloud runner is measured while
it runs. From the last runs of a job at its current size, the console suggests
one size up when the job is starved, or one size down when it leaves most of
the runner idle.

## What is measured

The runner samples the whole VM every 2 seconds. One VM runs one job, so the
numbers include any Docker containers the job starts. Only numbers are kept:

- CPU use (median, p95 and max, as a share of the runner's vCPUs),
- memory in use (median, p95 and peak, against the runner's total),
- IO wait (p95) and the bytes read from and written to disk,
- out-of-memory kills.

No command lines, process names, logs or file contents are collected.

Open a job in **Actions → Overview → History** to see the numbers of that run
next to the averages of the job's last 7 runs.

## When a size is suggested

Each job keeps a window of its last 7 runs at its current size. Cancelled runs,
runs shorter than 30 seconds and runs that were moved to a bigger runner are
left out.

| Suggestion | When |
| --- | --- |
| One size up | The newest run and another of the last 7 were killed for running out of memory, or in at least 5 of the last 7 runs CPU p95 was over 85% or memory peak was over 90% (at least 3 runs needed). |
| One size down | All of the last 7 runs had CPU p95 under 25% and memory peak under 40%. |

A suggestion clears once the job's runs no longer match it, or once the job's
`runs-on:` label is changed to the suggested size. After a suggestion in one
direction, none in the other direction is made for that job for 30 days.
Suggestions stay between 2 and 64 vCPU and never above your organization's
runner size limit. If the bigger size is above that limit, the suggestion says
**At your org's limit** and links to the **Limits** tab.

A smaller size shows the estimated saving per month, from the job's runs in the
last 30 days. A bigger size costs more per minute, but CPU-bound jobs often
finish fast enough that the total is similar.

## Applying a suggestion

Suggestions are listed under **Runner size suggestions** on the Actions
Overview tab. **Copy** puts the new label on the clipboard, for example:

```yaml
runs-on: a-line-4vcpu
```

Paste it into the job in your workflow file. The job starts a fresh window at
its new size. **Show runs** filters the History to the job's workflow.

## Local disk

Every runner has a 50 GB local disk, included in the per-minute price. A job
that needs more adds a disk part to its label, after the memory:

```yaml
runs-on: a-line-4vcpu-100gb
```

The size must be the full local disk of that runner size:

| vCPU | Extra storage |
| --- | --- |
| 2 | none (50 GB only) |
| 4 | 100 GB |
| 8 | 200 GB |
| 16 and up | 400 GB |

A label with any other size is rejected, and the error lists the valid
sizes. Leaving the part out keeps the included 50 GB. The setup tab under
**Actions** has a **Disk** choice that shows the price and builds the label
for you.

Extra storage is billed as **Actions extra storage**: the GB above 50, per
started hour of the job, at 0.0002 EUR per GB-hour by default. A 4 vCPU job
with 100 GB that runs 20 minutes pays 50 GB × 1 hour, 0.01 EUR, on top of its
compute. The jobs list shows the add-on next to each job's charge.

## Notifications

When a job gets a new suggestion, the organization's owner and admins get one
notification, at most once per job every 14 days and at most five a day for your organization, with memory
suggestions first; the rest follow later. Suggestions that are at your
organization's limit send no notification. Notifications follow the **Builds** switches in your
notification settings; turn those off to stop them. Members who can operate
the organization can also **Dismiss for 30 days** on a single suggestion.

## Cost and privacy

Measuring is free: it runs inside the job you already pay for, adds no runner
time and is not metered. The usage numbers and suggestions belong to your
organization only; no other organization can see them.
