Capacity and load tests
Measurements taken on 19 November 2026 on the Militant Visio production server: a 4-core VPS with 7.7 GiB of memory, running Coturn on the same machine. No human in the room, robots only.
Method
Section titled “Method”- Each robot is a headless Chromium driven by Playwright, with a fake camera and a fake microphone. All of them join the same room as
robot-NN. - A safeguard stops the test and removes every robot if the busiest mediasoup worker reaches 80 % of its core, if the VPS load reaches 3, or if free memory drops below 1 GiB.
- Each step is held for at least 30 seconds, with a reading every 5 to 13 seconds.
Version 1.6.5: one room, one worker
Section titled “Version 1.6.5: one room, one worker”Until version 1.7.0, a room ran on a single worker, so on a single core.
| People in the room | Worker CPU | VPS load | Free memory |
|---|---|---|---|
| 7 | 14 to 15 % | 1.0 | 4.8 GiB |
| 12 | 25 to 28 % | 0.8 | 4.8 GiB |
| 17 | 39 to 43 % | 1.0 | 4.8 GiB |
| 22 | 61 to 76 % | 1.6 to 1.9 | 4.7 GiB |
CPU grows by about 2.5 points per person, slightly faster as the room gets bigger. At 22 people the core reached 76 %; at that pace, 25 people would hit the 80 % threshold. Practical limit: 22 to 25 people with video per room. Coturn stayed under 1 %, because the robots connected directly.
Version 1.7.0: one room across three workers
Section titled “Version 1.7.0: one room across three workers”Since version 1.7.0, once a worker holds PEERS_PER_ROUTER people of a room (10 by default), the next ones go to another worker and streams are relayed between workers. The test used PEERS_PER_ROUTER=5 to see the spread with few robots.
| Robots | Spread per worker | CPU of each worker | Container CPU | VPS load |
|---|---|---|---|---|
| 5 | 5 / 0 / 0 | 11 to 13 % | 12 % | 1.1 to 1.8 |
| 10 | 5 / 5 / 0 | 14 to 16 % | 31 to 32 % | 1.0 to 1.1 |
| 15 | 5 / 5 / 5 | 15 to 23 % | 47 to 69 % | 1.1 to 1.3 |
| 15, microphones muted | 5 / 5 / 5 | 6 to 11 % | 18 to 47 % | 1.3 to 1.4 |
- The spread works as intended: 5 / 0 / 0, then 5 / 5 / 0, then 5 / 5 / 5.
- Compared with 1.6.5: 17 people on one worker took 39 to 43 % of its core. 15 robots on three workers take 15 to 23 % per core, so the busiest core carries about half.
- Relaying has a cost: total container CPU is higher than before at a comparable size, but it is spread over three cores.
- General assembly case, microphones muted: the CPU of each worker is roughly halved. Audio is heavy (microphone relaying and active-speaker detection).
Limits of the tests
Section titled “Limits of the tests”- The machine driving the robots saturated before the VPS did (load 42 to 74, memory at 90 %). The 15-robot figures and the muted step are therefore on the low side. The 20-robot step was not attempted.
- Robots connect directly, without a Coturn relay. People behind strict firewalls would load Coturn more.
- The measurements hold for a 4-core VPS; other machines will give other numbers.
Tuning
Section titled “Tuning”| Variable | Effect |
|---|---|
NUM_WORKERS |
Number of workers. Leave one core free for Coturn and the other services (for example 3 out of 4 cores). |
PEERS_PER_ROUTER |
Number of people of a room on one worker before the next ones go to another. 0 keeps a room on a single worker. |
HEALTH_DETAILS |
true: /health also reports the spread of people per worker. |
See also configuration.