Squid: "Socket error... (55) no buffer space available
 

Squid on the firewall machine sends the above error message when I want to get a webpage.
Netstat on the firewall showed me that there are hundreds of connections in TIME_WAIT state from port 3128 of the firewall machine to different ports of an other internal server of ours named "castor".

Netstat on "castor" does not show these connections. How is it possible?

I restarted our firewall machine (where squid is located), but the connections in TIME_WAIT state were there in some minutes, again.

What can I do to get rid of those unclosed connections?

Edit:

OK, problem solved: it was a download accelerator on a client behind "castor" that opened so many connections that they filled squid's buffer space.

The great experience was that such events are hardly traceable in our network topology: nat on castor hides (masquerades) the ip of the traitor, whilst netstat on castor does not seem to show connections forwarded by nat there. (So I had to run tcpdump on oif of castor to find the culprit).

Two hours later castor crashed. What the hell is going on here today?