The previous sections describe how the connection timeouts work when the once only flag is disabled on a cluster;
that is, when Equalizer is examining every set of headers received on a connection. The once only option, when
enabled, specifies that Equalizer will examine only the first set of headers received on a connection. This has the
following effects on connection timeouts:
l If you have once only enabled, as soon as the initial transaction (client request and server response) on a
connection completes, the connection goes into “streaming” mode and the client timeout is no longer used
for this connection. Equalizer does not parse any additional client requests received on the connection. The
server timeout is used for the remainder of the connection, and is reset whenever data is received from
either side of the connection.
l If you have once only disabled as described in the previous sections, and multiple requests are being sent
on the same connection, the client timeout starts counting down again as soon as a new request is received
from the client.
Layer 4 Connection Timeouts
Connections to Layer 4 clusters are received by Equalizer and forwarded with little processing. Equalizer simply
rewrites the source and/or the destination IP addresses, as appropriate for the cluster, and sends the packet to the
server specified by the cluster’s load balancing policy. For Layer 4 TCP clusters, a connection record is kept for
each connection so that address translation can be done on the packets going between the servers and clients.
The Layer 4 connection timeouts specify how long a connection record is kept by Equalizer.
Layer 4 TCP clusters use the idle timeout and stale timeout parameters. The idle timeout can be set at the global
and cluster levels, while stale timeout can be set at the global level only. The parameters affect how Equalizer
manages Layer 4 connection records:
l Connection records need to be removed in cases where the connection is not closed by the client or server,
and is left idle. If no data has been received on a connection from either the client or the server after the time
period specified by the idle timeout has elapsed, then Equalizer removes the connection record for that
connection. Any data received from either client or server resets the idle timer.
Note that when using Direct Server Return (DSR), the time that a connection record is maintained is
determined by adding the idle timeout for the cluster to the sticky time . This additional time is necessary
when using DSR, since no server responses are routed through Equalizer (and therefore cannot restart the
idle timeout to keep the connection open). For more information on DSR, see "Configuring Direct Server
Return" on page 313.
l In other cases, a connection may be initiated but never established, so the connection record goes “stale”
and must be removed. If a client fails to complete the TCP connection termination handshake sequence or
sends a SYN packet but does not respond to the server’s SYN/ACK, Equalizer marks the connection as
incomplete. The stale timeout is the length of time that a connection record for an incomplete connection is
maintained.
When Equalizer reclaims a connection, it sends a TCP RST (reset) packet to the server, enabling the server to
free any resources associated with the connection. (Equalizer does not send a TCP RST to the client when
reclaiming a connection.)
Reducing the stale timeout can be an effective way to counter the effects of SYN flood Denial of Service attacks
on server resources. A stale timeout of 10.0 (see table below) would be an appropriate value for a site under SYN
flood attack.
Copyright © 2013 Coyote Point Systems. A subsidiary of Fortinet, Inc.
All Rights Reserved.
265
Equalizer Administration Guide
Comentarios a estos manuales