---
title: Capacity and sizing
description: What one daemon holds, measured with the Computers load test. Memory per active browser, how many run at once, resume time, tool-call overhead, burst behaviour, and how to run the test on your own machine.
section: computers
order: 30
audience: [operator]
sources: [scripts/computers-load.mjs, scripts/computers-load-lib.mjs, docs/adr/0078-venom-computers.md, services/go-orchestrator/internal/computers]
features: []
---

These numbers come from `make computers-load`, which drives a real `venom-seatd` with Chrome over its HTTP and CDP API. The
run: 100 seats, 20 active for 3 minutes, a burst of 100 starts, 60 resumes, on an 8-core, 16-thread, 60 GiB machine that was
also doing other work.

| | Measured | Target |
|---|---|---|
| Seats provisioned | 100, no failures | at least 100 |
| Memory per active seat (PSS) | 194 MiB p50, 224 MiB p95 | at most 450 MiB p50 |
| Hibernated seat | no memory; 3.7 MiB of disk for a fresh profile | no memory, at most 300 MiB |
| Resume to first browser command | p50 199 ms, p95 333 ms | p50 1.5 s, p95 4 s |
| Tool call overhead over a local pipe | +0.11 ms p50 | at most +5 ms |
| Burst of 100 starts | all served in order, none failed | none fail, queued in order |
| CPU | 24% of one core for 20 active seats | |

## Sizing {#sizing}

Plan on **PSS**, the proportional memory the machine really pays, not RSS. A browser is about 14 processes that share pages,
so adding their RSS counts the same page many times (1.2 GiB per seat here, which is wrong as a cost).

At 194 MiB, a 16 GiB budget holds about 84 active seats. That is an extrapolation: the run kept 20 active. Real pages are
heavier than the loopback test site, so the daemon starts from 450 MiB per active seat when it computes `max_active` and then
follows what its hosts measure ([Venom Computers on a daemon](/docs/deploy/computers.md#admission)). Leave that default alone
unless you have measured your own pages.

Disk grows as agents browse (cookies, cache, downloads). The 3.7 MiB figure is a profile that has only opened a few pages;
watch the per-seat disk in the Computers app.

## Desktop computers {#desktop}

The same test with `--kind desktop` against the real image (ten active desktops for two minutes): 279 MiB per desktop at p50
and 293 MiB at p95, against 261 MiB for a browser in the same container, with a screenshot at 139 ms and a click at 63 ms. A
desktop therefore starts at 500 MiB in the daemon's budget ([desktop computers](/docs/computers/desktop.md#cost)). Inside a container the
daemon sizes by anonymous memory, since the proportional figure needs a privilege the host drops.

```bash
make computers-load ARGS="--kind desktop --docker venom-computer:dev --scenarios A,C --seats 10 --active 10 --duration 120"
```

## What the numbers do not cover {#limits}

- Pages are a local test site. Heavy single-page apps cost more memory and CPU.
- Screencast and thumbnail cost per viewer was measured on idle pages, which send few frames, so treat it as a lower bound.
- How often the daemon writes to the cloud, and the ten-minute idle hibernation, are daemon behaviour, not part of this test.
- The run is on one uid and without Docker. The container host adds a small start-up cost per host, not per seat.

## Run it on your machine {#run}

```bash
make computers-load ARGS="--quick"        # a minute: checks the script and your Chrome
make computers-load ARGS="--seats 100 --active 20 --duration 180 --max-active 10 --burst 100 --viewers 5"
```

It needs Go and Chrome or Chromium (`--chromium <path>`). It stops every seat and writes what it has if free memory falls under
`--min-avail-gib` (16 by default), so it cannot take the machine down. The report is written to `dist/computers-load/` as
Markdown and JSON.
