The primary objective of LAN connectivity in a cluster is to provide uninterrupted heartbeat communications. Use the guidelines in this section to design the LAN connectivity for the cluster:
Use a dedicated VLAN (virtual local area network) for each cluster.
The cluster protocol is non-routable, so you cannot direct communications to specific IP addresses. Using a VLAN for the cluster nodes provides a protected environment for the heartbeat process and ensures that heartbeat packets are exchanged only between the nodes of a given cluster.
When you use a VLAN, no foreign host can interfere with the heartbeat. For example, using a VLAN avoids broadcast storms that slow traffic and can result in false split-brain situations.
Servers should be redundantly cabled to the network in order to provide LAN fault tolerance, preferably at both the adapter level and the link level. Consider connecting cluster nodes to redundant access switches for fault tolerance.
Use channel bonding for the server adapters. Channel bonding combines Ethernet interfaces on a host computer for redundancy or increased throughput. Higher-level software uses a single virtual-network interface, and the channel bonding driver handles the complex choice of which physical-network interface to use. Channel bonding helps increase the availability of an individual cluster node, which helps avoid or reduce the occurrences of failover caused by slow LAN traffic. See the /usr/src/linux/Documentation/bonding.txt document.
Use the Spanning Tree Protocol (STP) to eliminate network topology loops. When you configure STP, ensure that the Portfast Bridge Protocol Data Unit (BPDU) guard feature is enabled, or consider using Rapid Spanning Tree Protocol (RSTP, IEEE 802.11w).
The default settings for STP inhibit the heartbeat for over 30 seconds whenever there is a change in link status. Test your STP configuration with Cluster Services running to ensure that a node is not cast out of the cluster when a broken link is restored.
Plan your IP address assignment so that it is consistently applied across each cluster. For each cluster, provide a dedicated IP address range with sufficient addresses for the cluster. The addresses do not need to be contiguous.
You need a unique static IP address for each of the following components of a cluster:
Cluster (master IP address)
Cluster nodes
Cluster resources (file system resources and service resources such as DHCP, DNS, SLP, FTP, and so on)
Ensure that SLP is properly configured for name resolution. See Section 4.6.4, SLP.