Load generation limits

A server only accepts connections, so every client shares its one listening port. A load generator is the other side: it opens every connection itself, and each one needs a source port of its own. That puts the generating machine up against a limit the server under test never sees, which matters as soon as a test needs many long-lived connections, such as thousands of WebSockets.

The ephemeral port limit

Each connection from the generator uses an ephemeral (temporary) source port. On a Mac (macOS) that range is 49152–65535, so one machine can hold only ~16k connections to one server address.

Phoenix’s 2M-WebSocket demo needed 45+ client machines for exactly this reason (see Connections per server).

Workarounds

To go past ~16k connections from one machine:

Have the server listen on several ports: each port is a separate destination with its own ~16k.
Use several loopback source addresses (127.0.0.2, 127.0.0.3 …), each with its own port range. On macOS each needs an alias:
Terminal window
```
sudo ifconfig lo0 alias 127.0.0.2
```
Raise the open-files limit on both sides: a Mac (macOS) caps one process at 92,160 (kern.maxfilesperproc).

A few thousand connections is enough to see reconnect storms, so these workarounds are rarely needed for that.