By default, the CloudBridge 4000/5000 Provisioning Wizard sets up load balancing to handle different kinds of connections appropriately. This default behavior is adequate for most installations.
Sending all the connections from the same remote accelerator to the same local accelerator maximizes the benefits of CloudBridge compression, and the default load balancing method accomplishes this. If an instance becomes overloaded or unavailable, new connections are reallocated.
By default, the NetScaler instance uses the least-connection method to balance the load across the accelerators. This method applies whether or not the connections are accelerated. Connections are persistent, but persistency is discontinued for an instance that becomes overloaded, and is lost if the local appliance is restarted or when no traffic from a remote appliance is seen for more than 24 hours.
For incoming accelerated connections (that is, connections with CloudBridge options in the header of the SYN packet), all connections from a given remote CloudBridge are sent to the same local accelerator.
The identity of the remote CloudBridge is determined by one of the CloudBridge SYN options: the "AgentID" field, which contains the management IP address of the remote CloudBridge.
This method is used for connections from remote CloudBridge appliances and remote CloudBridge Plug-ins.
Incoming non-accelerated connections and all outgoing connections are also distributed among the accelerators according to the least-connection method, but since they do not contain an AgentID field, they cannot use AgentID persistence. Instead, they use SRCIPDESTIP persistence, meaning that connections with the same IP addresses use the same accelerator.
If an instance is overloaded, the NetScaler instance bypasses it for new connections, sending them through without acceleration. Existing connections continue to be sent to the instance.
This behavior is controlled by the skipPersistency parameter. The default behavior is -skippersistency ReLB. The alternative behavior, -skippersistency bypass, instructs the NetScaler instance to pass the connection through without sending it to an accelerator.
The default load balancing behavior is adequate for most installations, but sometimes customization is needed. This is most commonly true when a few remote sites have much more traffic than the rest. In that case, it can be worthwhile to assign these large sites to accelerators explicitly.
Optional load balancing behavior includes the use of static routing (for hand-crafted load balancing) and variations on the least-connection with AgentID and SRCIPDESTIP persistence methods used in the default configuration. The behavior for dealing with overloaded instances can be changed from assigning connections to a difference instance to passing them through as unaccelerated.