Building a Windows Server 2008 Network Load Balancing Cluster
Windows Server 2008 supports both Network Load Balancing (NLB) and Failover clustering. This chapter covers Network Load Balancing clustering. The topic of Failover clustering is covered in detail in the chapter entitled Building a Windows Server 2008 Failover Cluster.
An Overview of Network Load Balancing Clusters
Network Load balancing provides failover and high scalability for Internet Protocol (IP) based services providing support for Transmission Control Protocol (TCP), User Datagram Protocol (UDP) and general Routing Encapsulation (GRE) traffic. Each server in a cluster is referred to as a node. Network Load Balance Clustering is supported by all versions of Windows Server 2008 with support for clusters containing 2 up to a maximum of 32 nodes.
Network Load Balancing assigns a virtual IP address to the cluster. When a client request arrives at the cluster this virtual IP address is mapped to the real address of a specific node in the cluster based on configuration settings and server availability. When a server fails, traffic is diverted to another server in the cluster. When the failed node is brought back online it is then re-assigned a share of the load. From a user perspective the load balanced cluster appears to to all intents and purposes as a single server represented by the virtual Ip address.
The failure of a node in a cluster is detected by the transmission of heartbeats by each node. If a node fails to transmit a heartbeat for a designated period of time that node is assumed to have failed and the remaining nodes takeover the work load of the failed server.
Nodes in a Network Load Balanced cluster typically do not share data, instead each storing a local copy of data. Under such a scenario the cluster is referred to as a farm. This approach is ideal for load balancing of web servers where the same static web site data is stored on each node. In an alternative configuring, referred to as a pack the nodes in the cluster all access shared data. In this scenario the data is partitioned such that each node in the cluster is responsible for accessing different parts of the shared data. This is commonly used with database servers, with each node having access to different parts of the database data with no overlap (a concept also known as shared nothing).
Network Load Balancing Models
Windows Server 2008 Network Load Balancing clustering can be configured using either one or two network adapters, although for maximum performance two adapters are recommended. In such a configuration one adapter is used for communication between cluster nodes (the cluster adaptor) and the other for communication with the outside network (the dedicated adapter).
The four basic Network Load Balancing modes are as follows:
- Unicast with Single Network Adapter - MAC address of network adapter is disabled and cluster MAC address is used. Traffic is received by all nodes in the cluster and filtered by NLB driver. Nodes in the cluster are able to communicate with addresses outside the cluster subnet but node to node communication within cluster subnet is not possible.
- Unicast with Multiple Network Adapters - MAC address of network adapter is disabled and cluster MAC address is used. Traffic is received by all nodes in the cluster and filtered by NLB driver. Nodes within the cluster are able to communicate with each other within the cluster subnet and also with addresses outside the subnet.
- Multicast with Single Network Adapters - Both network adapter and cluster MAC addresses are enabled. Nodes within the cluster are able to communicate with each other within the cluster subnet and also with addresses outside the subnet. Not recommended for significant levels of traffic are directed to specific cluster nodes.
- Multicast with Multiple Network Adapters - Both network adapter and cluster MAC addresses are enabled. Nodes within the cluster are able to communicate with each other within the cluster subnet and also with addresses outside the subnet. This is the ideal configuration for environments where there are significant levels of traffic directed to specific cluster nodes.
Configuring Port and Client Affinity
Network traffic arrives on one of a number of different ports (for example FTP traffic uses ports 20 and 21 while HTTP traffic uses port 80). Network Load Balancing may be configured on a port by port basis or range of ports. For each port three options are available to control the forwarding of the traffic:
- Single Host - Traffic to the designated port is forward to a singe node in the cluster.
- Multiple Hosts - Traffic to the designated port is distributed between the nodes in the cluster.
- Disabled - No filtering is performed.
Many client/server communications take place within a session. As such the server application will typically maintain some form of session sate during the client server transaction. Whilst this is not a problem in the case of a Single Host configuration described above, clearly problems may arise if a client is diverted to a different cluster node partway through a session since the new server will not have access to the session state. Windows Server 2008 Network Load Balancing addresses this issue by providing a number of client affinity configuration options. Client affinity involves the tracking of both destination port and source IP address information to optionally ensure that all traffic to a specific port from a client is directed to the same server in the cluster. The available Client affinity settings are as follows:
- Single - Requests from a single source IP address are directed to the same cluster node.
- Network - Requests originating from within the same Class C network address range are directed to the same cluster node.
- None - No client affinity. Requests are directed to nodes regardless of previous assignments.
