Our LVS/DR backends freezes

[Julian Anastasov]

	Hello,

On Tue, 28 Nov 2006, Olle Ö^Vstlund wrote:

> I read the postings about the ipvsadm connection-figures being erronous
> when using LVS/DR. These are figures of the director's ipvs-connections,
> taken from /proc/net/ip_vs_conn
> 
> STATE      CONNECTIONS
> -----------------------
> NONE            84 (max/mean expire time 59/33)
> FIN_WAIT     33581 (max/mean expire time 119/30)
> ESTABLISHED   1784 (max/mean expire time 900/201)
> CLOSE        38960 (max/mean expire time 59/29)
> SYN_RECV        49 (max/mean expire time 59/30)
> 
> It sure looks like a lot of non-terminted connections.

	Isn't that normal? Are the number of inactive connections
increasing hour by hour? 33581 conns in FIN_WAIT state can mean
average of 33581/120=279 conns destroyed per second. As the
established are 1784 may be one connection remains established
for 1784/279=6.39 seconds in average??? For example, send traffic
in 1.39 seconds and close after 5 seconds of web client inactivity?

	FIN_WAIT are terminated connections (from Client in the DR
case). This state exists to forward FIN retransmissions for DR/TUN and
for NAT to forward traffic from client when it is in CLOSE_WAIT/LAST_ACK 
state (after the RS sent FIN and is in FIN_WAIT state).

Regards

--
Julian Anastasov <ja at ssi.bg>