Understanding
Network Load Balancing
A Network Load Balancing
cluster consists of up to 32 servers, referred to as hosts, each of
which is running a duplicate copy of the application you want the
cluster to provide to clients. Network Load Balancing works by creating
on each host a virtual network adapter that represents the cluster as a single
entity. The virtual adapter has its own IP and media access control
(MAC) addresses, independent of the addresses assigned to the physical
network interface adapters in the computers. Clients address their
application requests to the cluster IP address, instead of an individual
server’s IP addresses.
Off the Record
In an Ethernet or
Token Ring network interface adapter, the MAC address, also known as
the adapter’s hardware address, is a unique six-byte hexadecimal value
hardcoded into the adapter by the manufacturer. Three bytes of the
address contain a code identifying the manufacturer, and three bytes
identify the adapter itself. |
|
Directing clients to
the IP address of the cluster is a task left to the name resolution
mechanism that provides clients with IP addresses. For example, if you
are currently running an individual Web server on the Internet, the DNS
server hosting your domain has a record associating your Web server’s
name with the Web server computer’s IP address. If you change from the
single Web server to a Network Load Balancing cluster to host your Web
site, you must modify the DNS resource record for the Web site’s name so
that it supplies clients with the cluster IP address, not your original
Web server’s IP address.
|
When an incoming client
request addressed to the cluster IP address arrives, all the hosts in
the cluster receive and process the message. On each host in an NLB
cluster, a Network Load Balancing service functions as a filter between
the cluster adapter and the computer’s TCP/IP stacks. This filter
enables NLB to calculate which host in the cluster should be responsible
for resolving the request. No communication between the hosts is
required for this purpose. Each host performs the same calculations
independently and decides whether it should process that request or not.
The algorithm the hosts use to perform these calculations changes only
when hosts are added to or removed from the cluster.
Planning a Network
Load Balancing Deployment
Before you deploy a
Network Load Balancing cluster, you must create a plan for the network
infrastructure that will support your cluster servers. The high
availability provided by NLB will do you no good if your users can’t
access the servers due to a failure in a router, switch, or Internet
connection. In addition, because many NLB installations provide Web and
other services to Internet users, you must consider the security of your
cluster servers and the rest of your internal network.
For a high-traffic Web site with
high availability requirements, a typical network infrastructure design
would consist of a Web server farm located on a perimeter network, as
shown in the following figure. The perimeter network has redundant
connections to the Internet, preferably with different Internet service
providers (ISPs) or with one ISP that has connections to multiple
Internet backbones. A firewall at each Internet access router protects
the perimeter network from Internet intruders, and another firewall
isolates the perimeter network from the internal network.
|
|
Important
Deploying a Network Load Balancing cluster is
not a task to undertake casually or haphazardly. As with any major
network service, the NLB deployment process must be planned carefully,
tested thoroughly on a lab network, and then implemented in a pilot
program before proceeding with the full production deployment. |
NLB Operational
Modes
The servers that are
going to be the hosts in your NLB cluster do not require any special
hardware. There is no shared data store as in a server cluster, for
example, so you do not have to build a storage area network. However,
NLB imposes certain limitations on a server with a single network
interface adapter in a standard configuration, and in some cases, you
can benefit from installing a second network interface adapter in each
of your servers.
Windows Server 2003
Network Load Balancing has two operational modes: unicast mode and multicast
mode. In unicast mode, Network Load
Balancing replaces the MAC address of the physical network interface
adapter in each server with the MAC address of the virtual adapter
representing the cluster. The server does not use the computer’s
original MAC address at all, effectively transforming the computer’s
physical network interface adapter into a virtual cluster adapter. The
Address Resolution Protocol (ARP) resolves both of the server’s IP
addresses (the IP address originally assigned to the network interface
adapter and the cluster IP address) to the single MAC address for the
cluster.
Off the Record
NLB does not
actually modify the MAC address in the network interface adapter itself;
the address assigned to the adapter by the manufacturer is permanent
and cannot be changed. NLB only replaces the MAC address in the
computer’s memory, substituting a virtual cluster address for the
physical address the system reads from the network adapter card. |
|
ARP is a TCP/IP
protocol that resolves IP addresses into MAC or hardware addresses. To
transmit to a particular IP address, a TCP/IP computer must first
discover the MAC address associated with that IP address, so that it can
build a datalink layer protocol frame. ARP functions by transmitting a
broadcast message containing an IP address to the local network. The
computer using that IP address is responsible for replying with a
message containing its MAC address.
In the case of an NLB
cluster in unicast mode, each server in the cluster replies to ARP
requests that contain either its original IP address or the cluster IP
address by sending a response containing the cluster MAC address.
Therefore, no computer on the network can transmit to the MAC address
assigned for NLB server’s physical network interface adapter.
|
Because the network interface adapters of all
the servers in the cluster have the same MAC address, the cluster
servers cannot communicate among themselves in the normal way, using
their individual MAC addresses. The servers can, however, communicate
with other computers on the same subnet, and with computers on other
subnets, as long as the IP datagrams don’t contain the cluster MAC
address.
Note
When you
configure the servers in an NLB cluster to use unicast mode with a
single network interface adapter, you cannot use the Network Load
Balancing Manager application on one of the servers to manage the other
servers in the cluster. |
In some cases, this is not
a problem. Dedicated Web servers hosting the same site, for example,
don’t often need to communicate with each other under normal conditions.
However, if you determine that it is necessary for the servers in your
NLB cluster to communicate with each other, there are two possible
solutions:
Configure the cluster
servers to operate in NLB multicast mode—
In multicast mode, NLB assigns a cluster MAC address to the physical
network interface adapter, but also retains the adapter’s original MAC
address. The cluster IP address resolves to the cluster MAC address and
the server’s original IP address resolves to the original MAC address.
For this configuration to function properly, the routers on the network
must support the use of multicast MAC addresses.
Install a second network interface adapter
in each server— One of the adapters
becomes the cluster adapter, with its original MAC address replaced by
the cluster MAC address. Both the cluster IP address and the adapter’s
original IP address resolve to the cluster MAC address. The system does
not use this adapter’s original MAC address. Like a single adapter in
unicast mode, the cluster adapter cannot communicate with the other
servers in the cluster. The second adapter retains its original MAC
address and assigned IP address and handles all noncluster network
communications.
Tip
In a
Windows Server 2003 Network Load Balancing cluster, you must configure
all the servers to operate in either unicast or multicast mode. You
cannot mix unicast and multicast servers in the same cluster. However,
you can mix network interface adapter configurations, installing two
network interface adapters in some of a cluster’s servers, while leaving
a single adapter in others. In the case of a unicast cluster, only the
servers with multiple adapters are able to communicate with the other
servers. |
In summary, a server in an
NLB cluster can have either one network interface adapter or multiple
adapters, and it can run in either unicast or multicast mode. By
combining these options, you can use four possible NLB configurations,
each of which has advantages and disadvantages, as shown in Table 1.
Table 1. NLB
Configuration Advantages and Disadvantages
| NLB Configuration | Advantages | Disadvantages |
|---|
| Single network interface
adapter in unicast mode | |
|
| Single network interface adapter
in multicast mode | |
|
| Multiple network interface adapters in unicast mode | No router incompatibility problems Permits ordinary communications among cluster servers Network performance enhanced,
because cluster traffic and ordinary network traffic use different
network interface adapters
|
|
| Multiple network interface adapters in multicast mode | Permits ordinary communications among cluster servers Network performance
enhanced, because cluster traffic and ordinary network traffic use
different network interface adapters
|
|
The most popular
configuration for large NLB installations is to install two network
interface adapters in each server and run them in unicast mode. This
enables the servers to function as normal participants on the network,
in addition to performing their NLB server duties. There are also no
problems with routers handling multicast MAC addresses and no
bottlenecks caused by cluster traffic and ordinary network traffic
sharing a single network interface adapter.
NLB Networking
Although the servers
in a Network Load Balancing cluster do not share a single data store,
as in a server cluster, and perform their own independent calculations
to determine which server will service an incoming request, the servers
do communicate with each other. The cluster servers must exchange
information to know many servers are in the cluster, and to determine
when a server has been added or removed from the cluster. This
communication enables the cluster to compensate for a failed server and
to take advantage of new servers in the cluster by redistributing the
traffic load.
Important
A single
computer running Windows Server 2003 cannot be a member of a Network
Load Balancing cluster and a server cluster at the same time, because
these two clustering solutions use network interface adapters in
different ways. If you want to deploy both an NLB cluster and a server
cluster on your network, you must use separate servers for each cluster. |
The cluster traffic
between NLB servers takes the form of a heartbeat message that each server transmits once per
second to the other servers in the cluster. If one cluster server fails,
it stops transmitting its heartbeat messages, and the other servers
detect the absence of the heartbeats. Once the other servers in the
cluster miss five consecutive heartbeat messages from a server, they
begin a process called convergence, in which they recalculate their traffic
distribution algorithm to compensate for the missing server. In the same
way, adding a new server to an NLB cluster introduces a new heartbeat
to the network, which triggers a convergence in the other servers,
enabling them to redistribute the traffic so that the new server
receives an equal share of the load.
Note
Because all the
servers in the cluster are using the same cluster MAC address,
transmitting the heartbeats is simply a matter of directing the packets
to that address. The servers don’t need to broadcast the heartbeat
messages, reducing the impact of the cluster traffic on the network. |
When you deploy NLB
cluster servers with a single network interface adapter in each
computer, obviously all the cluster-related traffic must travel over the
same network as your ordinary traffic. This is usually not a major
burden, because the heartbeat packets are small, less than 1,500 bytes,
so they fit into a single Ethernet packet. If you decide to install
multiple network interface adapters in each cluster server, you can
connect both adapters to the same local area network (LAN) or construct a
separate network for the cluster traffic.
Planning
If your NLB
cluster consists of servers that are already isolated on a perimeter
network, there is probably no need to create a separate LAN for cluster
traffic. However, if you are deploying an NLB cluster on a heavily
trafficked internal network, you might benefit from installing a
dedicated cluster LAN. |
