As Layer 3 switching has grown in popularity, it has demonstrated that ATM is not the only technology capable of great speed. However, ATM does have its place in many campus networks. This section examines some of the more important issues associated with completing an ATM-based campus network design.
When to Use ATM
One of the first questions every network designer must face is should the design utilize ATM technology. In the past, ATM has been billed as the solution to every possible network problem. Although this might be true in terms of ATM’s theoretical capabilities, it is not true in terms of how most organizations are using ATM. For example, in the mid-1990s, many network analysts foretold of the coming days where networks would use ATM on an end-to-end basis. Instead, Ethernet has continued to grow in popularity. When, then, is it best to use cell-based switching?
Traditionally, ATM has been touted for several unique benefits. The most commonly mentioned benefits include:
- High bandwidth— Because cells use fixed-size units of data with simple and predictable header formats, it is fairly easy to create high-speed hardware-based switching equipment.
- Sophisticated bandwidth sharing— Cells can be interleaved to allow multiple communication sessions to share a single link through an advanced form of statistical multiplexing. Because cells are all the same size, applications using large data transfer units do not create a log jam effect that slows down smaller and potentially more time-sensitive traffic.
- Quality of Service (QoS)— ATM has complex and sophisticated mechanisms to allow detailed traffic contracts to be specified and enforced.
- Support for voice and video— ATM’s low latency and QoS benefit give it robust support for time-critical forms of communication such as voice and video.
- Distance— Unlike common campus technologies such as Ethernet, ATM can function over any distance.
- Interoperability— Because ATM is a global standard, a wide variety of devices can be purchased from different vendors.
Although many of these points remain true, advances in frame-based switching have significantly eroded ATMs edge in the following areas:
- Campus-oriented Gigabit Ethernet switches now match or exceed the speed of ATM switches. Although cell switching does maintain a theoretical advantage, ASIC-based Layer 2 and Layer 3 switches have become exceptionally fast. Furthermore, ATM has continued to struggle with the SAR boundary, the fastest speed that ATM’s Segmentation And Reassembly function can be performed.
- Ethernet-based QoS (or at least Class of Service) schemes are becoming more available, more practical, and more effective. Although ATM holds a theoretical lead, ATM continues to suffer from a lack of applications that capitalize on its inherently superior capabilities. As a result, CoS-capable Gigabit Ethernet switches are rapidly growing in popularity.
- Although ATM does maintain a distinct advantage in its capability to handle isochronous (timing critical) applications, there is tremendous growth in non-isochronous mechanisms for sending voice and video traffic. Efforts such as voice over IP (VoIP) and H.323 videoconferencing are common examples. These technologies reduce the need for ATM’s unique capabilities.
- Gigabit Ethernet distances are growing dramatically. As this book goes to press, a number of vendors are introducing 80–100 km Gigabit Ethernet products.
- All forms of Ethernet, including Gigabit Ethernet, have been perceived as being considerably more interoperable than ATM standards.
In addition, the complexity of ATM has become a significant issue for most organizations. Whereas Ethernet is considered easy and familiar, ATM is considered difficult and murky (and, to a significant extent, these perceptions are valid).
Although the growth of ATM in campus networks has slowed at the time this book goes to press, it is important to note that the use of ATM technology in the WAN continues to expand rapidly.
Where to Use ATM
Although there is considerable debate about the usefulness of ATM in a campus backbone, there is considerably less debate about where it is useful. Almost all analysts are in agreement that desktop connections will be Ethernet for the foreseeable future. Although 10/100 Ethernet sales continue to soar, sales of ATM to the desktop have staggered. When ATM is used, almost all agree that the ATM is best suited to the core of the network. In most cases, this means a LANE core connecting to Ethernet switches containing LANE uplink modules.
Although this issue has received fairly little debate, a second issue has been less clear-cut. The issue concerns the matter of how far the ATM backbone should reach. The debate surrounds two options.
Some vendors and network designers prefer to link only the MDF/distribution layer devices to the ATM core. Fast and Gigabit Ethernet links can then be used to connect to IDF switches as shown in Figure 15-12.
Figure 15-12. Using Ethernet Links in Conjunction with an ATM Core
The advantage of this approach is that it uses cost-effective Ethernet technology in the potentially large number of IDF closets. This design is often deployed using the campus-wide VLAN model to extend the speed of ATM through the Ethernet links. The downside is that it creates a large number of Layer 2 loops where redundant MDF-to-IDF links are used. Unfortunately, these links have been shown to create Spanning Tree loops that can disable the entire campus network. Furthermore, it is harder to use ATM features such as QoS when the edges of the network use Ethernet.
The opposing view is that the ATM backbone should extend all the way to the IDF closets. Under this design, the entire network utilizes ATM except for the links that directly connect to end-user devices. This approach is illustrated in Figure 15-13.
Figure 15-13. Extending the ATM Core to the IDF Switches
The downside of this alternative is a potentially higher cost because it requires more ATM uplink and switch ports. However, the major benefit of this design is that it eliminates the Layer 2 loops formed by the Ethernet links in the previous approach. Because LANE inherently creates a loop-free Layer 2 topology, the risk of Spanning Tree problems is considerably less (in fact, some vendors who promote this design leave Spanning Tree disabled by default, something many network engineers feel is a risky move).
Having worked with implementations using both designs, I feel that the answer should be driven by the use of Layer 3 switching (like many other things). If you are using the multilayer model to create hard Layer 3 barriers in the MDF/distribution layer devices, the MDF switches can be the attachment point to the ATM core and Ethernet links to the IDF devices can be safely used. However, when the campus-wide VLAN model is in use, extending the ATM backbone to the IDFs allows for the most stable and scalable design. Trying to use the MDF-attachment method with campus-wide VLANs results in Spanning Tree loops and network stability issues.
The use of Layer 3 switching in your network should drive the design of an ATM core.
Until standards-based LANE redundancy mechanisms become widely available, Simple Server Redundancy Protocol (SSRP) will remain an important feature in almost any LANE-based core using Cisco ATM switches. Although SSRP allows more than one set of redundant devices, experience has shown that this can lead to scaling problems. See Chapter 9 for more information on SSRP.
Always try to place your LANE Broadcast and Unknown Server (BUS) on a Catalyst LANE module. Because the BUS must handle every broadcast and multicast packet in the ELAN (at least in current versions of the protocols), the potential traffic volume can be extremely high. The Catalyst 5000 OC-3 and Catalyst 5000/6000 OC-12 LANE modules offer approximately 130 kpps and 450 kpps of BUS performance respectively, considerably more than any other Cisco device currently offered.
One decision faced by designers of large LANE cores involves whether a single BUS or multiple distributed BUSes should be utilized. The advantage of a single BUS is that every ELAN has the same logical topology (at least the primary topologies are the same, the backup SSRP topology is obviously different). The disadvantage is that the single BUS can more easily become a bottleneck.
Distributed BUSes allow each ELAN to have a different BUS. Although this can offer significantly higher aggregate BUS throughput, it can make the network harder to manage and troubleshoot. With the introduction of OC-12 LANE modules and their extremely high BUS performance, it is generally advisable to use a single BUS and capitalize on the simplicity of having a single logical topology for every ELAN.
With the high BUS throughput available with modern equipment, centralized BUS designs are most common today.
Chapter 9 contains additional information on BUS placement.
Multiprotocol Over ATM (MPOA) can be a useful technology for improving Layer 3 performance. MPOA, as discussed in Chapter 10, “Trunking with Multiprotocol over ATM,” allows shortcut virtual circuits to be created and avoids the use of routers for extended flows. When considering the use of MPOA, keep the following points in mind:
- MPOA can only create shortcuts in sections of the network that use ATM. Therefore, if the MDF devices attach to an ATM core but Ethernet is used to connect from the MDF to the IDF switches, MPOA is only useful within the core itself. If the core does not contain Layer 3 hops, MPOA offers no advantage over LANE. In general, MPOA is most useful when the ATM cloud extends to the IDF/access layer switches.
- Because MPOA is mainly designed for networks using ATM on an IDF-to-IDF basis, you must intentionally build Layer 3 barriers into the network. Without careful planning, MPOA can lead to flat earth networks and the associated scaling problems discussed earlier in this chapter and in Chapters 11, 14, and 17.
- At presstime, significant questions remain about the stability and scalability of MPOA.
MPOA only optimizes unicast traffic (however, related protocols such as a MARS can be used to improve multicast performance).
In most Catalyst equipment (such as the Catalyst 5000), both MPOA and LANE use MAC addresses from the chassis or Supervisor to automatically generate ATM NSAP addresses. For a detailed discussion of how NSAP addresses are created, refer to Chapter 9. When designing an ATM network, keep the following address-related points in mind:
- Devices with active backplanes such as the Catalyst 5500s use MAC addresses pulled from the backplane itself. Changing the chassis of one of these devices therefore changes the automatically-generated NSAP addresses.
- Devices with passive backplanes such as the Catalyst 5000 use MAC addresses from the Supervisor. Therefore, changing a Catalyst 5000 Supervisor module changes the pool of addresses used for automatically generating NSAP addresses.
- In both cases, 16 MAC addresses are assigned to each slot. Therefore, simply moving a LANE module to a different slot alters the automatically generated NSAP addresses.
- Because of these concerns, many organizations prefer to use hard-coded NSAP addresses. For more information, see the section “Using Hard-Coded Addresses” in Chapter 9.
Consider using hard-coded NSAP addresses in a large LANE network.