Reduce or eliminate subnets and routers to increase your network's speed
In "Techniques to Speed Up Your NT Network" (April 1997), Joel
Sloss describes how to create a network to connect clients to Windows NT
servers. In this article, I'll show you another network design that can increase
network speeds even more. I always try to design networks that install easily,
work fast, reduce costs, and increase reliability. To create such networks, you
need to avoid subnets and reduce the number of routers.
To Subnet or Not to Subnet: That's the Real Question
One major difference between Joel's design and mine is the use of IP
subnets. Joel's design features four subnets. I prefer to configure a network
without subnets because their use increases complexity and usually decreases
performance.
Ethernet switching provides an easy way to segment a network without
subnets. The difference between Ethernet switching and IP routing is how deep
the device must look into the packet to determine where to send it. In IP
routing, the device must look deep into the packet. In Ethernet switching, the
device has to look at only the first few bytes, enabling much faster traversal
of the network. Ethernet switches are protocol independent and require minimal,
if any, configuration. In other words, they're almost Plug and Play.
You can also use IP switches (new devices that offer the speeds of switches
and the ability to route packets simultaneously), but their cost is high. As the
price of IP switches drops, they will become viable for networks. But at this
time, you can justify them for only extremely high-speed TCP/IP networks.
In addition to speed and simplicity, switches offer versatility. With
Ethernet switches, you can mix 10Base-T and 100Base-T adapters on the same
network, decreasing costs. Many Ethernet adapters feature either 10 Megabits per
second (Mbps) or 100Mbps operation. The only difference is the port they plug
into. Because the network automatically detects the adapter's speed, you don't
need any protocol or setup modifications.
With Ethernet adapters, you can connect 10Base-T hubs to 10/100 network
cards in your workstations and servers. Then as the need develops, you simply
plug the 10/100 network card into a 100Base-T hub, increasing throughput without
touching the workstation.
Routers: A Path Not Taken?
Another difference between Joel's design and mine is how the network
connects to other corporate networks or the Internet. Joel suggests that you
create a router-to-router connection to link networks. Routers, however, have
too many configuration parameters that you can inadvertently misconfigure. As a
result, I suggest that you use a port on the existing corporate router to link
to other corporate networks. To connect to the Internet, you probably need
another router to link to your Internet Service Provider (ISP). You might also
need a firewall to provide protection.
| Table 1: Cost Per Port
|
| Port Type | Cost |
| 10Base-T Hub | $10 |
| 100Base-T Hub | $60 |
| 10Base-T Switched | $100 |
| 100Base-T Switched | $600 |
| 10Base-T to 100Base-T Converter | $500 |
To minimize costs while keeping network throughput high, I use 10Base-T
ports for standard workstations, 100Base-T ports for advanced workstations, and
100Base-T switched ports for servers. As
Table 1 shows, ports for 10Base-T are
inexpensive. But, if you have the money, you can connect everyone to 100Base-T,
which will increase throughput for network intensive applications.
Although I try to avoid using routers and subnets, they do have their
place. I can't always design a network without subnets. Subnets work effectively
for linking locations with low-speed connections, connecting large numbers of
computers, and setting up networks that have many protocols.
Similarly, you might need to use routers. But router configuration isn't
for the beginner. With a few hundred parameters to configure, you need to know
what you are doing.
Putting Thoughts to Paper: The Network Design
With these considerations in mind, you can start putting the network design
on paper. To begin, you must determine the configuration of the hubs, switches,
and routers. An average small office needs to provide for about 50 low-speed hub
connections, 12 high-speed connections, and 6 high-speed switched connections
for the servers.
You can use low-speed connections for a majority of the devices on the
network, such as standard workstations, printers, routers, and other
instruments. You need to use high-speed connections for engineering workstations
and other devices that need high-speed access but usually talk to only one or
two other devices. You must use high-speed switched connections to servers or to
any device that needs high-speed access and connects to many different devices
on your network.
You can configure the network in many different ways. I like to use a
10/100 switch as the central point and connect the 10Base-T hubs, 100Base-T
hubs, and servers to it. Figure 1 shows this configuration.
The amount of available bandwidth can help you determine whether to use a
hub, switch, or router as the central point in your network. As Figure 2 shows,
using a hub is like using one garden hose to connect all the devices. All
information flows through the same line.
Using a switch or router is like using several garden hoses to connect each
device on the switch or router. Because information flows through several lines,
throughput increases significantly.
Routers and switches differ in their total speed capability. Routing
imparts a significant penalty: Many smaller routers have problems keeping up
with 10Base-T. Some smaller routers can't even handle a T-1 line at 1.544Mbps.
Most switches provide full throughput between ports, providing a total of
400Mbps for an 8-port 100Base-T switch.
To provide the fastest connections to the application servers, I give the
servers a dedicated 100Mbps port on the switch. If the switch and the network
adapter support full duplex, you can run both devices at 200Mbps with no
collisions. This configuration provides an extremely fast connection from the
clients to the server. Collisions are isolated to the user segments, and each
server can obtain full 200Mbps throughput.
Designing the Network Services
After designing the network, you need to create the network services design.
Because you don't have multiple subnets to worry about, you might be tempted to
put all the network services--such as Primary Domain Controller (PDC), Domain
Name System (DNS), Windows Internet Name Service (WINS), and Dynamic Host
Configuration Protocol (DHCP)--on one server. But you don't want to build in a
single point of failure.
Instead, you can use two servers, each capable of performing all the
necessary services. Neither server needs to be very large or fast, just
reliable. Two 486 or small Pentium systems, for example, can easily fill the
needs of up to 100 users.
When designing the services, you first need to determine the domain
controller architecture. An NT domain controller provides security for a
network. It lets you centralize user administration to provide a fairly secure
network. The NT domain system consists of a PDC and any number of Backup Domain
Controllers (BDCs).
Although BDCs are optional, I strongly suggest having at least one. If your
PDC fails and you don't have a BDC, you will lose all security information and
the ability to access most of the network. Thus, you need to configure both a
PDC and BDC. (For more information about how to configure PDCs and BDCs, see Ed
Tittel and Mary Madden, "PDCs, BDCs, and Availability," August 1996.)