With the introduction of the proprietary StackWise interconnect technology that ties its new line of Catalyst 3750 switches together, Cisco Systems has filled the gap between its small, stand-alone Ethernet switches and its large, modular chassis and blade switches.
The Catalyst 3750 stack is a good fit for enterprise applications where a large chassis solution is not economically feasible, and smaller stand-alone switch products don’t scale. That’s what we concluded after we connected three Catalyst 3750s in a ring — you can connect up to nine boxes — and found that the ring provided up to 30Gbit/sec throughput while maintaining low latency, condensed administration and management features, and increased redundancy.
Cisco offers four switch configurations in the Catalyst 3750 line — a 24-port 10/100Base-T switch with two small form-factor pluggable (SFP) uplinks, a 48-port 10/100Base-T switch with four SFP ports, a 24-port 10/100/1000Base-T switch and 24-port 10/100/1000Base-T switch with four SFP ports.
We tested the 24-port Fast Ethernet box tied to both 24-port Gigabit switches.
Rack and stack
The switches are interconnected using proprietary multi-pin connectors. Two of these connections per switch are needed to form a ring. The switch hardware load-balances all packets entering the switch onto both directions of the stack ring.
Even if a packet is destined for a port on the same switch it entered, it will be forwarded to the ring. The ring acts as the backplane for all the switches in a stack. In a full-ring configuration, Cisco says it can support 32Gbit/sec of throughput.
If a stack cable or switch fails, the neighbouring switches will sense the ring disconnect and terminate the ring on both sides of the fault. The ring can support 16Gbit/sec in this fault state, Cisco says.
Our tests showed that ring throughput is 26Gbit/sec for the smallest allowable packet size and 30Gbit/sec for the largest packet size. This difference in bandwidth is most likely because of the extra header information tacked onto each packet entering the ring and the bandwidth used by the token to arbitrate access to the ring. When we introduced a fault in the ring by unplugging one of the ring cables, the ring bandwidth was measured to be about half of the full ring bandwidth, or 15Gbit/sec with large packets and 13Gbit/sec with small packets.
We also tested the ring configuration for increased latency. The Catalyst 3750 stack racked up a worst-case latency of 50 microsec, which is well below the threshold where latency-sensitive applications would be affected.
When you power up a Catalyst 3750 stack, a master switch is selected for the ring according to six criteria documented by Cisco. The first criterion is a user-defined master election priority. The ability to prioritise the master election process worked as expected. We defined which switch would always become master, which switch would become master if it fails and so on until one switch is left in the stack.
Stack and switch configuration and ongoing management are handled through two interfaces: a command-line interface (CLI) or cluster management services (CMS).
The CLI looks and feels like the standard Cisco interface with the addition of a few commands to configure stack options. The stack appears as a multiple card chassis with the typical hub/slot/port numbering convention. The master switch replicates the CLI to all the switches in the stack so that console port access shows the same information from any switch in the stack. The configuration files are unified for all switches. If the master switch fails, the new elected master will have the configuration for the entire stack.
To make sure all switches can implement all features known to the master switch, the master copies its software image to each switch in the stack if necessary. This removes the risk of a switch not having to handle a particular feature configured through the master switch. Because this copying process might infringe on software licensing, the administrator must make sure the proper number of software licences has been purchased.
To unify the switch configuration, each switch in the stack is given an arbitrary ID number. This number is used when displaying the interface ports. We tested the ability to renumber the switch IDs. This worked as expected and made administrative tasks much easier.
The Java-based CMS supports many browsers, but it did not work on an Apple PowerBook running OS X.
CMS has two modes of operation: expert and guide. Expert mode basically replicates the configuration options found in the CLI. Guide mode provides configuration wizards that help configure the stack for complex functions such as quality of service and multicast routing. The wizards were useful but limited in flexibility.
One beneficial stack management feature is remote deployment. Once a stack is up and running on the network, any technician should be able to connect a new switch into the stack and power it up. At that point, the network administrator can configure the new switch without any local network configuration and remotely manage it.
While the performance, features and flexibility offered by the Catalyst 3750 bundle stack up very well, we pinpointed two troublesome spots. The master election and configuration unification processes do not handle multiple ring faults well. The stack system was designed to only handle one fault at a time. If two ring faults occur, the stack becomes two stacks each with a master of its own. After reconnecting the ring without a reboot, the master election process might not be able to resolve a new master. You can reboot the stack to remedy this issue. This shortcoming doesn’t negate the usefulness of the architecture, but it’s something to plan for if you deploy these switches in your network.
Because the switches have to keep up with things like master election, CLI replication, unifying the configuration, and other basic stack functions, the time it takes to reboot the switches is longer than you might be used to. A switch with no stack connections will reboot in about two minutes. A three-switch stack takes about two minutes and 15 seconds to reboot.
One very useful feature of the stack is cross-channel etherchannel configuration. This feature lets you configure multiple links to belong to the same logical link or bundle. These links do not have to belong to the same switch. Without the stack feature, link aggregation is restricted to ports on the same switch. This feature worked as advertised.
John Bass is a senior technical staff member at North Carolina State University’s Centennial Networking Labs and a member of the Network World Global Test Alliance