Determining the Network Infrastructure Requirements
Applies To: Windows Server 2003, Windows Server 2003 R2, Windows Server 2003 with SP1, Windows Server 2003 with SP2
The network infrastructure affects your Network Load Balancing solution more than any other component. No matter how much Network Load Balancing enables you to scale out your solution, an inadequate routing and switching infrastructure can create a number of problems.
Even if you optimize the Network Load Balancing cluster, an inadequate routing and switching infrastructure can restrict available network throughput and prevent clients from achieving any improvement in response times. Additionally, an inadequate routing and switching infrastructure can allow a single failure in a router, switch, or network path to disrupt communications with the cluster.
Determine the network infrastructure requirements for your cluster by completing the following steps:
Determine the IP subnet requirements for the cluster.
Determine how the cluster handles inbound and outbound traffic.
Determine when to include switches or hubs to connect cluster hosts to one another.
For more information about how the network infrastructure affects availability, see "Ensuring Availability in NLB Solutions" later in this chapter. For more information about how the network infrastructure affects scalability, see "Scaling NLB Solutions" later in this chapter.
Determining Cluster IP Subnet Requirements
Network Load Balancing requires that all cluster hosts be on the same IP subnet or virtual IP subnet (VLAN). This is because all cluster hosts share the cluster’s virtual IP address. The routing infrastructure sends client requests to all cluster hosts by using a virtualized unicast MAC address or a multicast MAC address. For more information about unicast and multicast addressing, see "Selecting the Unicast or Multicast Method of Distributing Incoming Requests" earlier in this chapter.
Determining How the Cluster Handles Inbound and Outbound Traffic
Network Load Balancing cluster traffic is handled differently for inbound and outbound traffic. The differences between inbound client requests and outbound responses affect the network infrastructure and the use of switches and hubs.
Inbound cluster traffic is sent to all cluster hosts by using broadcasts or multicast traffic, which is required so that all cluster hosts can receive the inbound traffic. Switches learn where to send packets by watching the source packets that are sent from the computers directly connected to the switch. After the switch learns a computer’s location, the switch sends subsequent packets to the same switch port.
From an inbound-traffic perspective, the behavior of the switch is similar to a hub, because inbound cluster traffic is sent to all switch ports. However, from an outbound-traffic perspective, the switch provides isolation by preventing other cluster hosts from seeing a response to a client from the cluster host that services the client request.
For more information about the behavior of Network Load Balancing traffic, see the Network Load Balancing Technical Overview link on the Web Resources page at https://www.microsoft.com/windows/reskits/webresources.
Determining When to Include Switches or Hubs for Interconnecting Cluster Hosts
For most networks, switches are the preferred technology for connecting network devices to one another. Although many existing network infrastructures have hubs, switches are typically used in new deployments.
Although Network Load Balancing works in most configurations of switches and hubs, some configurations allow for optimal performance, and they provide ease of maintenance and operations. The recommended configuration is to use switches for interconnecting cluster hosts. Configurations that include hubs are also supported, but they require more complex network infrastructure and configuration.
The supported configurations for switches and hubs include:
Cluster hosts connected to a switch that is dedicated to the cluster.
Cluster hosts connected to a switch that is shared with other devices.
Cluster hosts connected to a hub that is connected to a switch.
Cluster hosts connected to a switch that is dedicated to the cluster
In this configuration, the cluster hosts connect to a switch that is dedicated to the cluster. Because the inbound cluster traffic is sent to all ports on the switch, the primary advantage of a switch for inbound traffic — segregating traffic to a limited number of switch ports — is lost.
However, for outbound traffic, only the cluster host responding to the client request is aware of the traffic, because the switch isolates the outbound traffic to the port connected to the responding cluster host. This reduces the congestion of traffic for all the cluster hosts that are connected to the switch.
Cluster hosts connected to a switch that is shared with other devices
In this configuration, the cluster hosts are connected to a switch that is shared with other devices. The switch sends inbound cluster traffic to the other devices as well, creating unnecessary traffic for the other devices.
You can isolate the inbound cluster traffic to only the cluster hosts by establishing a VLAN comprising the ports that connect to the cluster hosts. After you establish the VLAN, inbound cluster traffic is sent only to the cluster hosts and not to the other devices that are attached to the same switch. Using a VLAN to segregate inbound client traffic works for the unicast or multicast method of distributing incoming requests. You can also use the multicast method with IGMP to limit inbound cluster traffic only to the cluster hosts.
Because outbound cluster traffic is sent directly to the client that originated the request, only the cluster host responding to the request sees the outbound cluster traffic. All other cluster hosts, and the other devices connected to the same switch, are unaware of the outbound cluster traffic.
Cluster hosts connected to a hub that is connected to a switch
From the perspective of inbound cluster traffic, a switch provides no additional benefit over a hub, except for being able to define a VLAN that isolates inbound traffic from other devices attached to the switch. Because all cluster hosts receive the inbound cluster traffic, the switch sends all inbound traffic to all ports that are attached to the switch (or ports that are assigned to the same VLAN).
From the perspective of outbound cluster traffic, a switch provides a greater advantage, because it reduces the network contention for outbound cluster traffic. With a switch, only the cluster host originating the outbound cluster traffic is aware of the traffic, because outbound traffic is switched. Using a hub causes network contention for all cluster hosts. With a hub, all cluster hosts receive, and subsequently determine if they need to process, the outbound cluster traffic.
To eliminate the network contention caused by using a hub, install an additional network adapter in each cluster host for the purpose of responding to client traffic. Inbound requests are received through the cluster adapters in each cluster host, while the outbound responses are sent to the clients through the additional network adapter.
Because of the increased complexity of configuration (adding an additional network adapter to each cluster host) and network infrastructure (adding an additional switch to connect all the additional network adapters to the clients), this is not a recommended configuration. Because a switch is required for the additional network adapters, it is easier to eliminate the additional network adapters and connect the cluster adapters to a switch.
If you are considering adding network adapters to improve cluster performance, consider increasing the data rate of the cluster network adapters and corresponding network infrastructure, instead of adding network adapters. For example, to increase available network bandwidth, specify 100 megabits per second (Mbps) network adapters instead of 10 Mbps network adapters (along with the appropriate upgrades to intermediary switches and routers). For more information about increasing the available network bandwidth to the cluster, see "Increasing Available Network Bandwidth to the Cluster" later in this chapter.