Speeds and feeds are not all that one needs. Indeed, there are many features and factors to consider when shopping for a backbone switch.
Are you concerned with interoperability? Standards should be tops on your list. What about quality of service (QoS)? Queuing, congestion control and traffic prioritization should be your priority. And how can you manage without management? The switch's ability to troubleshoot should not be a trouble spot.
These are just some of the features you'll find listed in our online
Buyer's Guide chart, which includes 15 backbone switches from 13 vendors. We limited the chart to modular, multilayer (Layer 2 and 3) switches capable of being configured with at least 96 10/100M bit/sec Ethernet or 100Base-FX ports and a minimum of four Gigabit Ethernet ports. The exception is Plaintree's WaveSwitch 9200, which has been shipping for over a year with 10/100M interfaces but without the requisite Gigabit Ethernet ports, which are slated to ship after we go to press.
You might notice the absence of some other prominent vendors. We didn't include NEO Networks' SteamProcessor 2400 switch because it's still in beta testing. Also missing from the chart is FORE Systems, which acquired Berkeley Networks in late August. FORE opted not to list its eligible switch in our chart because the company plans to relaunch Berkeley's switches under the FORE brand name this fall.
Standard fare
Each vendor claims to have a feature-rich backbone switch. Some features are givens - like a rear window defogger on your car, they should be there without asking for them. Others are options vendors use to differentiate their offerings and to entice you to pay a little bit more for their switches.
For starters, every Layer 3 Ethernet switch should support a routing protocol, the 802.1D Spanning Tree algorithm and SNMP management. Every switch in our chart checks out here: Each one supports at least Routing Information Protocol (RIP), RIP2 and Open Shortest Path First Version 2, as well as Spanning Tree and SNMP.
Additionally, support for wire-speed multicasting on all ports is common to all entries, and each switch is flexible in terms of management. All can be managed by either command line or Web browser, and most have their own graphical management application as well.
Next, let's consider hot-swappable components, which are becoming standard fare for users that need to make upgrades without taking down their networks. This is where we start to see some differentiation. Eight vendors claim that all of the components of their switches - switching fabrics, interface cards, power supplies and fans - are hot swappable.
The remaining five vendors fall short of supplying total hot-swappable components. Compaq, Foundry, Lucent and ODS Networks do not offer hot-swappable switch fabrics, nor does Compaq offer hot-swappable power supplies or a cooling system on its Digital GIGA-switch/Router. Extreme does not offer hot-swappable interface modules on its Black Diamond switch, which is interesting because interface modules are usually the first switch components to be swapped.
Redundancy is another popular, if not vital, feature for backbone switches. If a backbone switch goes down and there's no backup, a whole company can be paralyzed until that switch is brought back up or replaced.
With redundancy, we again see divergence in vendor offerings. Only six of the 13 vendors claim redundancy for all critical switching components. Nortel Networks' Bay Networks division offers redundant power supplies and interface modules, but interestingly, not a redundant switching fabric for the Accelar switch. Next to power supplies, switching fabrics are arguably the most important switch component to back up.
Foundry's BigIron 8000 also fails in the switch fabric failover department.
Looking at hot swappability and redundancy, the differences in the degree of network uptime vendors can offer are clear. If network uptime - or downtime, for that matter - matters most in your network purchase decisions, you can already weed out some of the products that don't back up your backbone.
QoS is another important consideration for users as traffic increases and takes on less deterministic characteristics with the advent of the Internet and the World Wide Web. To ensure that the right people are receiving the right information at the right time, users should size up the amount of buffer memory and number of per-port queues, and the types of policies switches can support.
With regard to buffer memory, vendors are all over the map. The low is 256K bytes per port for Lucent's P550 Cajun and XLNT's Millennium 4000+ switches, while the high is Packet Engines' PowerRail 2200, which scales from 14M to 40M bytes per port. (Alcatel last week announced it will purchase Packet Engines.)
Per-port queues are fairly consistent from vendor to vendor, ranging from two to eight, with four being about the mean. Then we come to Cisco, with an astronomical 32,000 priority queues per port. Now that's granularity. Are there really 32,000 levels of priority you can assign a packet or flow? Buyer, beware.
For policy, virtually all vendors support Layer 2 and Layer 3, and application-level policy establishment and enforcement. All comply with the latest standards, including the IEEE's 802.1p priority queuing scheme and 802.1Q virtual LAN protocol. But the ability to allocate a minimum or maximum level of bandwidth per group or on an end-to-end basis is spotty.
Check the numbers
Users should be wary of vendor backplane capacity numbers because some clever computational techniques are being practiced.
We asked all the vendors in our chart to double-check their throughput numbers, some of which were suspiciously high. We wondered how 3Com, for instance, could get 560G bit/sec out of the CoreBuilder 9000 backplane when the ATM switching fabric only scales from 15G to 70G bit/sec. 3Com admitted that analysts prodded them to play with packet trace, redundancy and full-duplex figures to come up with a true monster of an enterprise switch. Then they cut the number in half - to 280G bit/sec, a number we still question - for our chart.
3Com isn't the only vendor that submitted revised figures. Cisco used Cisco math to originally claim 50G bit/sec out of the Catalyst 5500's 3.6G bit/sec Supervisor III switching engine. This seemed high, especially since Cisco user Eastman Kodak told us it has to devise ways of keeping Gigabit Ethernet traffic off the Catalyst backplane so performance doesn't bog down (NW, Sept. 14, page 74). Cisco later lowered its number to 9G bit/sec for our chart.
In general, it would be wise to question vendors that claim a throughput greater than 100 million packet/sec, as Foundry originally did. Analysts and consultants say each Gigabit Ethernet port can deliver a maximum of 1.5 million packet/sec. So in order to deliver 100 million packet/sec, Foundry, for example, needs to be able to populate its box with 86 Gigabit Ethernet ports. But the company lists its maximum Gigabit Ethernet port density at 64 ports. When we asked vendors to double-check their figures, Foundry lowered its original number from 100 million to 27 million packet/sec maximum aggregate wire-speed throughput.
Lastly, users should nail down vendors on pricing. Per-port costs are often quoted for Layer 2 ports; make sure you consider that Layer 3 10/100M ports are typically $100 to $150 more than Layer 2 ports.
So let the online chart be the start, not a final determinant, of your backbone switch purchase. Use it to pin vendors down on their claims and have them spell out exactly how they arrive at the numbers they did. Their wizardry may astound you.
Duffy is a senior editor with Network World. He can be reached at
jduffy@nww.com.
RELATED LINKS
