What has two names, four slots, switches or routes at 32G bit/sec, and leaves most of the other Gigabit Ethernet switches we've tested in the dust?
Hewlett-Packard calls it the ProCurve Routing Switch 9304M, and that's the version we evaluated. But Foundry Networks, which actually makes the HP hardware, calls it the BigIron 4000. Either way, with each of its four chassis slots filled with eight-port gigabit modules, the switch delivered maximum traffic loads on each port and didn't miss a beat (or a bit).
We first tested Layer 2 media access control (MAC) switching with static streams. That's where all of port 1's traffic is switched to port 32, all of port 2's input is switched to port 31, and so on. Then we tested Layer 2 switching using the "mesh test" feature of Ixia Communications' IXIA 1600 traffic generator/ analyzer. This is a grueling workout, in which packets arriving at a gigabit per second on each port (that's 1.488 million minimum-sized packets per second) are sequentially switched to each of the other ports in round-robin fashion. Picture a big round table, at which each of 32 dealers is dealing cards to the others at a rate of 1.488 million cards per second.
Under all Layer 2 configurations we found 100% wire-speed performance. Next we tried Layer 3 IP routing, after setting up IP subnets on each of the 32 ports.
We got the same results: 100% wire-speed performance, even for round-robin, mesh test routing among all ports.
In virtually every other performance metric, the switch checks out just fine.
Head-of-line blocking? There wasn't any. Latency? It ranged from a mere 7 microseconds for 64-byte packets under a light switch load to 70 microseconds with a full load on all ports. Jitter? The worst we saw was 17 microseconds with a full load on all ports. Those numbers indicate there's nothing to worry about with latency or jitter. And our tests showed there's no leakage between virtual LANs (VLAN).
The only unusual behavior we noted related to the switch's quality-of-service (QoS) features. The switch lets you assign as many as eight priorities to different traffic classes, based on criteria including MAC address, protocol, IP source or destination address, TCP or User Data-gram Protocol (UDP) port numbers, or ranges of port numbers. It turns out, though, that these are mapped to only four queues, so you might say that the switch really supports only four discrete classes of traffic.
We assigned a top QoS priority to all UDP traffic, as a user might do to make sure that voice-over-IP traffic was always expedited through the network ahead of other, non-time-sensitive data. The HP switch dutifully gave priority to UDP traffic, but only up to a maximum of 80% of available bandwidth. We checked with the vendor, who told us there's a weighted-fair-queuing algorithm built into the switch's logic, which always reserves some bandwidth - 20% as it turns out - for other traffic sources and types.
We really can't argue with that. If you've got more than 80% of your bandwidth consumed by your top priority traffic, you might want to think of rearranging your priorities.
The switch's management functions are vested mainly in HP's browser-based Java interface, a feature not included with Foundry's version of the switch. The application is straightforward and fairly intuitive, but lacks any online help.
Without that, you have to regularly stop and fumble through the documentation.
The written documentation was clear and legible, with a full table of contents but an anemic index.
The application displays traffic statistics as raw numbers - a graphical display would be better. And while the statistics are updated automatically, someone forgot to put the units of measure on the display. You don't know whether you're looking at bits, bytes or packets. And because there is no on-screen help, you can only guess.
You can limit broadcast and multicast traffic switchwide, but we feel that per-port broadcast control would be better. Generally speaking, though, traffic flow controls, including per-port filtering, are adequate.
The command-line interface has a Cisco feel in its command structure and syntax, and overall is fairly useful. But the switch's management isn't helped much by HP's TopTools - a set of general-purpose management applications, which HP ships free with the switch. While graphical, the Windows-based TopTools suite is not that intuitive to use. It features a discovery capability, which adroitly turned up a couple of test tools we didn't think were speaking on our lab network. But other than that, TopTools didn't seem particularly useful.
As it did for performance and management, the switch fared well in the category of scalability and fault tolerance. All of the same modules in this model also work in HP's eight-slot model of the switch, the ProCurve 9308M, allowing for a graceful upgrade path.
Redundant power supplies and fans are available, and all modules are hot-swappable. We found a small problem with the redundant management modules, however. We tried out a new version of the vendor's redundant management cards, but we couldn't get automatic fail-over to work. HP acknowledged that the redundant management, still in late beta when we tested it, wasn't working properly and told us the company was addressing this. New, higher-priced redundant management modules are required to implement this feature.
Under the broad heading of functionality we rated this switch fairly high; as a Layer 2 switch or Layer 3 IP router there isn't much missing. Among the features are VLANs, QoS, and full IP routing - including Border Gateway Protocol 4, Routing Information Protocol Versions 1 and 2, and Open Shortest Path First, as well as the set of new IP multicasting protocols.
Multilink trunking is supported. As many as four gigabit links can be aggregated in a logical trunk group between two switches or between, say, a switch and a server with multiple network interface cards. You can assign traffic over a multilink group in a couple of ways to achieve some degree of load balancing. This setup also provides link failover protection. We tested it and it works; traffic from a failed link is shifted over to a working link in about 0.3 seconds - actually a little faster than some competitors' gigabit switches.
We tested with the vendor's 1000BaseSx port modules, but the switch also supports longer-distance 1000BaseLx, plus a full range of Ethernet and Fast Ethernet modules with as many as 24 ports per module. It lacks WAN interfaces such as OC-3, OC-12 or DS-3. And gigabit-over-copper - four-pair unshielded twisted pair, per the newly adopted IEEE standard - won't be available until next year.
The bottom line? HP's ProCurve 9304M is a performance dynamo, and its broad functionality and good scalability features make it a recommended buy.
(Mier (firstname.lastname@example.org) is founder and president of Mier Communications (http://www.mier.com), a network consultancy and product test center based in Princeton Junction, New Jersey. Smithers (email@example.com) is vice president, technology and Maksymow (firstname.lastname@example.org) is manager of lab testing.)