Previous Table of Contents Next


Implement Features at the Appropriate Layer

This guideline is one of the most difficult to enforce, yet it is one of the most important. Included in this policy is the recommendation that access lists remain outside of the core layer. While Cisco has greatly improved the performance of their router products, access lists and other services still impose a substantial burden on resources (depending on router type and features). By keeping these functions at a deeper layer of the model, the designer should be able to maintain performance for the majority of packets. Each design will require some interpretation of this guideline—there clearly may be exceptions where a feature must be deployed at a specific point in the network.

Network Design Issues

All good network designs will address at least one of the following questions. Excellent designs will answer all of them:

  What problem are we trying to solve?
  What future needs do we anticipate?
  What is the projected lifespan of this network?

What Problem?

New networks are typically deployed to solve a business problem. Since there is no legacy network, there are few issues regarding the existing infrastructure to address. Existing networks confronted by a potential upgrade are typically designed to resolve at least one of the problems discussed below, under “Considerations of Network Design.”

Future Needs?

It is unlikely that anyone with the ability to accurately predict the future would use such ability to design networks. Ignorance is a likely enemy of efforts to add longevity to the network design. An assessment of future needs will incorporate a number of areas that will help augment the lifespan of the network, but success is frequently found in “gut feelings” and overspending.

Network Lifespan?

Many would classify this topic as part of the future needs assessment; however, it should be viewed as a separate component. The lifespan of the network should also not be viewed in terms of a single span of time. For example, copper and fiber installations should be planned with at least a 10-year horizon, whereas network core devices that remain static for more than 36 months are rare. Given these variations, it is important to balance the costs of each network component with the likelihood that it will be replaced quickly. Building in expandability and upgradability will affect the lifespan of a network installed today. Designers should always consider how they might expand their designs to accommodate additional users or services before committing to a strategy.

Considerations of Network Design

The network design considerations addressed in this section are the solutions to the network design issues addressed earlier. For example, the first network design consideration below addresses excessive broadcasts. The designer will need to understand the concept of broadcasts in the network, how they are impacting the existing network, how they may increase in the future, and how broadcasts may be dealt with in the lifespan of the network.

Excessive Broadcasts

Recall that broadcasts are used in networking to dispatch a packet to all stations on the network. This may be in the form of an Address Resolution Protocol (ARP) query or a NetBIOS name query, for example. All stations will listen and accept broadcast packets for processing by an upper-layer process—the broadcast itself is a Layer 2 process.

While the broadcast packet is no larger than any other packet on the media, it is received by all stations. This results in every station halting the local process to address the packet that has been forwarded from the network interface card. This added processing is very inefficient and, for the majority of stations, unnecessary.

A general network design guideline says that 100 broadcasts per second will reduce the available CPU on a Pentium 90 processor by two percent. Note that this figure does not compare the percentage of broadcasts on the network to user data (typically unicast). While most modern networks are now using much more powerful processors and larger amounts of bandwidth per workstation, broadcasts are still an area warranting control by the network designer and administrator.

There are two methods for controlling broadcasts in the network. Routers control the broadcast domain. Thus, a router could be used to divide a single network into two smaller ones. This would theoretically reduce the number of broadcasts per segment by 50 percent. This technique would also affect bandwidth and media contention, so it might be the correct solution. However, it’s now much easier to use a router to reduce broadcasts. In reality, the total number of broadcasts will almost always increase when using two networks instead of one. This is due to the nature of the upper-layer protocols. For example, a single network could use a single Service Advertising Protocol (SAP) packet (Novell), whereas a dual network installation will require at least two. The number of broadcasts per network will decrease, but not by 50 percent.

Another method for controlling broadcasts is to remove them at the source—typically servers and, to a lesser extent, workstations. This is one aspect of network design that greatly benefits from the designer having a detailed knowledge of both protocols and operating systems. For example, Apple computer has offered an IP-based solution for its traditional Apple-Talk networks for a long time. Implementation of this service would greatly reduce the number of broadcasts in the network for a number of reasons, including the elimination of an entire protocol and AppleTalk’s intensive use of broadcasts. Assuming that most workstations are also running IP for Internet connectivity, this design could easily be incorporated into the network. Removing AppleTalk provides two benefits—a reduction in background broadcasts compared with IP and in the amount of overhead demanded by the network.

Contention for the Media

Media contention is frequently associated with 10Mbps Ethernet, where a large number of stations are waiting for access to the physical layer and a large number of collisions are likely to occur. However, media contention can also occur in FDDI and Token Ring. While both of these technologies negate the possibility of collision, each station must wait for receipt of the token before transmitting. This can cause significant delays.

Historically, access to the media was controlled by installing additional router ports and hubs. Installing new routers may result in network-wide IP readdressing, which may have a large up-front cost factor. While installing these routers reduced the number of stations on the segment, it did not eliminate contention issues; rather, it reduced the impact and frequency of them. With the advent of switching technology in the network, designers were offered the opportunity to virtually eliminate contention at a low cost. Discounting buffering issues and other advanced considerations, a full-duplex connection presents no contention points. This is a marked improvement that may be implemented with no change to the user workstation (with the possible exception of a full-duplex-capable network card). Designers should consider the use of switching technologies to resolve media-contention issues.


Previous Table of Contents Next