Most users assume that their networks are always available. However, most networks have periods of downtime quite regularly. Network administrators need to reprogram switches and routers and perform other maintenance operations; they do this at times of low network usage and the outages last for only a few seconds. A cluster uses it's interconnect continuously, 24 hours a day, 7 days a week. As a result, a routine network maintenance operation that would not even be noticed by a typical network user or application can result in a cluster recovery action that shuts down systems and moves applications.
Temporary outages can result from other causes as well. One of the most convenient improvements to Ethernet has been the speed-agile ports on hubs, switches, and routers. These allow the user to plug an Ethernet device into a port and not have to worry about speed matches between the network interface card (NIC) and the port. However, there is a cost to this technology. If a hub or switch gets too busy, it can lose its state and will have to renegotiate with the NIC. This renegotiation takes several seconds, during which time no messages are passed. As a result, a cluster interconnect on this type of device could disconnect long enough so that the cluster takes a recovery action such as shutting down a node.
Another temporary outage can occur when an Ethernet device is first opened. This starts the chain of events that triggers the switch port to which the Ethernet device is attached to begin its configuration process. Some switches take over one minute for this configuration to complete. PRIMECLUSTER may interpret the lack of responses on this port incorrectly and assume no other cluster nodes are configured on this port.
See
Because different network configurations and devices have different requirements, PRIMECLUSTER allows the system administrator to tune the parameters associated with the cluster interconnect. Refer to "PRIMECLUSTER Cluster Foundation (CF) Configuration and Administration Guide" for more details.
