Network Load Balancing

Network Load Balancing with Network Hardware Switches

As explained in Windows 2000 Network Load Balancing Help, network adapters for Network Load Balancing hosts in a single cluster must all be on the same broadcast subnet and connected to each other through either a hub or a switch.

Network hardware switches mediate between a network and computers or other switches, routing packets from the network to the correct computer.

If you connect Network Load Balancing hosts with a switch, the switch must be level-2 rather than level-3 or higher, because all the hosts share the same IP address (the cluster IP address), and level-3 switches direct network packets (incoming client requests) according to the IP address of the destination computer.

In unicast mode, each host's unique media access control address is replaced with the same cluster media access control address. Identifying all the hosts with one media access control address makes it possible to distribute incoming client requests (network packets) to all the hosts.

However, most level-2 switches require that each port be associated with a unique source media access control address. Network Load Balancing addresses this requirement in unicast mode by enabling the MaskSourceMAC feature by default.

When MaskSourceMAC is enabled, Network Load Balancing masks the source media access control address for outgoing packets so that for each port the switch continues to see a unique source media access control address. This satisfies the switch's requirement that each port be associated with a unique media access control address. Figure 19.1 shows a representative configuration of a Network Load Balancing cluster in unicast mode, with MaskSourceMAC enabled, and attached to a level-2 switch.

Enlarge figure

Figure 19.1 Network Load Balancing cluster running in unicast mode with MaskSourceMAC enabled

Masking the cluster media access control address on outgoing packets prevents the switch from associating the cluster media access control address with a single port. When a client request (which contains the cluster media access control address) enters the switch, the switch does not recognize the media access control address in the packet and so sends the packet to all ports. This is called "switch flooding."

In unicast mode, Network Load Balancing induces switch flooding by design, so that packets sent to the cluster's virtual IP address go to all the cluster hosts. Switch flooding is part of the Network Load Balancing strategy of obtaining the best throughput for any specific load of client requests.

If, however, the cluster shares the switch with other (noncluster) computers or other clusters, switch flooding can add to the other computers' network overhead by including them in the flooding.

You can avoid flooding noncluster computers by putting a network hub between the switch and the Network Load Balancing cluster hosts, and then disabling the MaskSourceMAC feature. The hub delivers each packet to every host, and the switch associates the cluster media access control address with a single port, satisfying the switch's requirement that each port be associated with a unique media access control address.

Placing the Network Load Balancing hosts downstream (toward the cluster) from a hub does not reduce the bandwidth for downstream packets. However, all upstream (from the cluster) traffic must flow through the hub. To optimize use of the hub, you can also connect each host's second network adapter back to another port in the switch for outbound packets, as shown in Figure 19.2. This has the following benefits:

Routing outbound packets through network adapters that are not attached to the hub improves use of the hub's capacity.
Use of the capacity for multiple upstream pipes from the switch to the network is improved, because multiple cluster hosts can simultaneously send traffic to different upstream pipes.
Using two network adapters to separate each cluster host's inbound and outbound network traffic improves the cluster hosts' handling of network traffic.

Enlarge figure

Figure 19.2 Network Load Balancing cluster running in unicast mode, with MaskSourceMAC disabled

Finally, if you choose not to use a hub as described here (for example, if the Network Load Balancing cluster does not share the level-2 switch with any other computers), you can put a level-3 switch upstream from the level-2 switch to prevent switch flooding of other interconnected level-2 switches.