Re: neighbors/neighborUp

Justin Pietsch wrote:

> Anyways, code was added in two places. The first in
> neighborsUdpPing: if there have been no requests for more than
> Config.deadPeer seconds then a host that would be marked down is given a
> fake last_reply time. The second is in neighborUp where a host is counted
> as up if there have been no requests for more than Config.deadPeer
> seconds.

I have two patches which remakes this in a more (in my opiniton) clean
manner.

ICP:
* Peers are initially UP
* If no ICP replies to any ICP query has been received within a
dead_peer_timeout interval then the peer is marked as down.
* Peers marked as down are probed with at most one ICP query each
dead_peer_timeout interval by including it in a ICP peer selection round
(no reply expected, but it will be used if one is). The peer stays
marked as down until a ICP reply is received.

TCP:
* Peers are initially UP
* When 10 TCP connections in a row has failed to a peer, the peer is
marked as down.
* A dead TCP peer is probed with a TCP connection at most each
connect_timeout interval. The peer stays marked as down until this probe
is successful in establishing a TCP connection.

Both ICP and TCP probing above is traffic driven. If there is no traffic
then dead peers won't be probed.

The patches are available from my Squid page
<http://hem.passagen.se/hno/squid/>.

--
Henrik Nordstrom
Spare time Squid hacker