HP StorageWorks Scalable File Share System Installation and Upgrade Guide Version 2.2

Preparing for the installation2–14
2.4.2.2.1 Symmetric dual Gigabit Ethernet network configuration
In a symmetric configuration, the number of interconnect interfaces per server and the number of
interconnect interfaces per client node are the same. In this case, the simplest way of configuring the
networks is to use a separate network for each link and to ensure that the networks cannot
intercommunicate.
Dual-link server to dual-link client nodes configuration
Figure 2-7 shows a very simple setup for a symmetrical configuration.
Figure 2-7 Dual-link server to dual-link client nodes — networks physically separate
In this setup, the two logical networks are also physically separate networks, although this is not a
requirement. The configuration shown here satisfies two of the requirements listed in Section 2.4.2.2: the
interconnect interfaces are on separate logical networks, and the path between interfaces on the same
network traverses as little network hardware as possible. The third requirement (that identically addressed
non-interconnect networks on client nodes must be excluded from use when mounting file systems) is only
an issue if the network interface that is not intended to connect to the file system has an IP address on either
the 10.128.0.0 or the 10.128.8.0 network; if this is the case, the network must be excluded from use when
mounting file systems.
The configuration in Figure 2-7 also satisfies the best practice guidelines listed in Section 2.4.2.2, as the
networks and netmasks configured on intercommunicating interfaces are identical on the servers and client
nodes, the physical topology is optimal, and there are no routers needed for the client nodes and servers
to communicate.
If the two networks are not physically separate, the arp_ignore parameter on all interconnect interfaces
on the client nodes must be set to 1; this is good practice in all cases where a client has more than one
interconnect interface, but is essential when the two networks are not physically separate. The default
behavior of the Linux kernel is to respond to an ARP request on any interface when the IP address on the
request is configured on any interface on the client node. Such behavior can lead to traffic for both networks
being received by a single interface on the client, effectively halving the read bandwidth of a single client
node. Setting the arp_ignore parameter to 1 stipulates that an interface receiving an ARP request can
only respond to the request if the IP address of the request is an address assigned to the receiving interface.