3.3. Network Topology

There are only two main network pathways used by the cluster: the frontside, or public, and the backside, or private, cluster interconnect network.

Clients or application servers mostly use the public network in order to connect to the database. When a node fails, existing transactions, sessions and connections disappear and this can create an interruption in service to these connections. The decision to deploy Cold Failover or RAC/GFS depends on how fast connections and transactions must restart. Cold Failover does not preserve any state, but can still restart very quickly, without having to reconstruct, re-connect, and re-synchronize with the application. RAC provides the ability to preserve much more context about sessions and transactions. If configured properly (including the application tier), this can dramatically reduce the downtime, but it increases both cost and complexity.

The most difficult situation is with existing connections that have opened a TCP/IP socket to the database. When the database node fails, the client socket needs to be notified as soon as possible. Most JDBC drivers now use out-of-band signaling to avoid the dreaded hung socket. Connection pools within application servers must be configured correctly, so failover delay is minimized.

The backside network is a private, dedicated network that should be configured as a four-port VLAN, if a non-private switch is used.

Most customers buy dual-ported NICs, which are not as reliable as two single-ported NICs. However, bonding ports across different drivers is also not recommended (bonding a TG3 port and an e1000 port, for instance). If possible use twp outboard single-ported NICs. Servers with the same out-board ports as the built-in ports (all e1000 ports, for instance), can safely cross-bond.

Connecting the ports to two different switches may also not work in some cases, so creating a fully redundant bonding NIC pathway is harder than it should be. Since the goal of the back-side network is for heartbeat, if the NIC fails but the server is up the server is still fenced. Statistically, the cluster might fence a little more often, but that’s about it.