sysctl.si › Parameters › net.core.somaxconn

net.core.somaxconn

Longest queue of finished connections waiting for accept().

Default
4096 (128 before 5.4)
Typical value
4096 to 65535

When to change it

Raise it for busy listeners such as nginx or HAProxy. The application's own backlog setting must also be raised to benefit.

How it works

When a server calls listen(), the kernel keeps a queue of connections that have finished the TCP handshake but have not yet been accepted by the application. net.core.somaxconn is the largest queue length the kernel allows.

The effective queue is the smaller of this value and the backlog the application asks for. Kernels before 5.4 default to 128, and newer ones default to 4096.

Signs you need to change it

Recommended values by workload

Small service, light traffic
4096 (default on modern kernels)
Busy web server or proxy
16384
Very high connection rates
65535

Common mistakes

Check the current value

sysctl net.core.somaxconn

Change it until the next reboot

sudo sysctl -w net.core.somaxconn=4096

Make it permanent

echo "net.core.somaxconn = 4096" | sudo tee /etc/sysctl.d/99-custom.conf
sudo sysctl --system

Related parameters

net.ipv4.tcp_max_syn_backlog · net.core.netdev_max_backlog · net.core.rmem_max · net.core.wmem_max · net.ipv4.tcp_congestion_control · net.core.default_qdisc · net.ipv4.tcp_tw_reuse · net.ipv4.tcp_fin_timeout

Generate a full sysctl config for your server

Values are starting points. Test on a staging machine and measure before and after.