Conversation

Beware. Each time you call *setsockopt(SO_SNDBUF)* even with a high value, you are likely _decreasing_ your max send buffer! The default max for linux tcp autotune is 4MiB. By calling setsockopt() you are _disabling_ autotune! The default net.core.wmem_max is just 208KiB.

2

7

25

Replying to

Where the autotune is configured? How to disable it system-wide? I have a limited amount of RAM and several tens of thousands of connections.

1

Replying to

and

Oh, you're talking about net.core.wmem_default/net.core.wmem_max/net.ipv4.tcp_wmem

1

Replying to

and

You can save a lot of memory without actually losing throughput by setting net.ipv4.tcp_notsent_lowat to a value like 128k to cap how much buffer space it allocates beyond the scaled window size. As long as your applications can fill the buffer quickly, you won't lose throughput.

6:29 PM · Jun 8, 2022Twitter Web App