Subject: Re: How to get rid of client disconnected network sockets [changing
tcp_keep_alive timeouts]
Am 12.06.25 um 13:24 schrieb Andrzej Adam Filip:
AFAIR Classic recommendation is to change tcp keep_alive timeout.
It changes when kernel checks (sends probe packets) over inactive
tcp connection. The default is after 2 hours (7200 seconds).
AFAIR some people reported changing it to 10 or 15 minutes.
I played a bit around with different settings, but with only limited
success.
The problem is primarily with excessive retransmissions. The ping
latency becomes up to 20 seconds (!) while packet loss is quite low.
I got some improvement by significantly reducing the TX queue depth of
the WLAN device (50 instead of 1000). It makes no sense to send a packet
20 seconds later.
The basic problem seems to be that the transmission power is restricted
to 15 dBm although channel 100 allows significantly higher level (26
dBm). The router seems to be aware of that but the Intel Centrino N 6205 device not. So link quality is highly asymmetric.
No idea whether the power limit is due to hardware restrictions or a
driver bug. I guess the first.
However, I found a workaround for the orphaned connections. Using systemd-socket-proxyd I can control almost all socket options very fime-grained. It also starts the VM if not yet running.
Marcel
--- MBSE BBS v1.1.1 (Linux-x86_64)
* Origin: MB-NET.NET for Open-News-Network e.V. (3:633/280.2@fidonet)