Make the workgroup switch connection

On a shared 10-Mbps hub, each client computer gets only about 3 Mbps of bandwidth on
average. If all ports are busy, the number drops because each client’s network
interface card is trying to get a free line through the hub.

Contrast that congestion with that of a switched network.

Each switch port receives dedicated bandwidth, similar to a dedicated phone line. A
server that needs more bandwidth and handles high-priority traffic can connect directly to
the switch. PCs used for videoconferencing also can attach directly to the switch to
ensure reliable high-speed bandwidth. Individual users who need bandwidth less
consistently can hook up to a hub that in turn connects to the switch

The GCN Lab looked at five 24-port workgroup switches that have auto-sensing
10/100-Mbps ports because the dual rate support eases the upgrade path for PCs. The lab
evaluated the switches for performance, manageability, scalability, upgradeability and
value. The extensive tests relied on the SmartBits 2000 device from NetCom Systems of
Chatsworth, Calif. (see story this page).

The switches came from 3Com Corp., Allied Telesyn International Inc., Cisco Systems
Inc., Compaq Computer Corp. and Hewlett-Packard Co. Any one of the switches could improve
an existing shared-hub network architecture. But I wanted to see how well they would
perform under heavy use, how easy they were to manage and what options they offered for
growth.

Many of the switches could stack together with others from the same makers. The lab did
not test this function, but it gave them extra points for scalability.

An excellent switch with 24 ports costs about the same today as a 24-port hub did a few
years ago. The range of prices is fairly surprising—from about $1,300 for the HP
switch to more than $5,000 for the switch released in March by Allied Telesyn.

I was surprised by how many of the makers differed from the company names on the
labels. The Compaq SW3322, for example, in fact came from NetVantage Inc., recently
acquired by Cabletron Systems Inc. of Rochester, N.H.

The SW3322 arrived with Compaq manuals and Compaq-branded management software. But
Allied Telesyn’s switch did not seem to know where it came from. Not only was its box
labeled by Extreme Networks of Cupertino, Calif., but the Telnet-accessible operating
system and embedded Web server documentation mentioned only Extreme, not Allied Telesyn.
Extreme Networks’ Summit 24 switch looks identical, except for its purple color, to
the Allied switch I tested.

The lab awarded the Reviewer’s Choice designation to the Hewlett-Packard ProCurve
2424M. Not only was it one of the top performers, its price was far lower than the
others’. That should come as no surprise to readers who attended the FOSE ’99
trade show in Washington where the HP 2424M won the Best New Product Award in the
communication products category.

The 2424M strikes an excellent balance of performance, manageability, scalability and
price. It can stack with another 2424M switch without losing any of its 24 10/100-Mbps
ports.

The boring but necessary console menus were easy to navigate through, as on many other
HP network devices. Best of all, an embedded Web server made it possible to manage the
2424M. Some of the other switches had embedded Web servers, but the ProCurve’s
interface was easier and more reliable.

An expansion slot for Gigabit Ethernet or fiber 100-Mbps connections, as well as good
scalability and manageability, suit the 2424M to high-use wiring closets and to offices
looking to grow at a low price.

The 3Com SuperStack II Switch 3300 looked good on paper because of its extensive
connectivity uplink options, stacking matrix option and future support for Layer 3
switching. But lackluster performance on some of the lab’s tests plus a shortfall in
manageability put the switch in the No. 4 position in our roundup.

Most switches have either a dedicated memory size per port or a shared memory pool.
Both approaches have advantages, depending on the type and amount of network traffic. The
SuperStack II 3300 combines the approaches with 32K dedicated memory per port as well as a
2.5M shared memory pool. This should ensure at least average performance in all
situations, but our measurements revealed poor throughput and many lost packets.

Perhaps in a stacked switching configuration the 3300 switch would perform better. But
on a standalone basis, it was at best average. Its advantages are stacking features,
extensive connectivity options, future upgradeability to Layer 3 switching and low price.

Offices that are looking for high port density, media compatibility and an upgrade path
to Layer 3 switching media will find good value in the 3300. But be prepared to cope with
difficult management software.

The 3300 is available on the open market for almost $800 less than on General Services
Administration Information Technology schedule contracts.

In spite of its identity crisis, Allied Telesyn’s CentreCom 8525SX-L2 switch did
well in the lab’s performance tests and had several excellent high-end features.

Two Gigabit Ethernet uplink ports, one of which acts as a redundant link, will fit into
many demanding environments. The CentreCom would likely have beaten the HP ProCurve except
for its much higher price.

Extreme Networks announced its version of the switch in November at the same price
Allied Telesyn is charging now. The performance and scalability features are certainly
best-of-breed, but the CentreCom has not yet appeared on GSA schedule contracts, and its
$5,000-plus price compared with other 24-port 10/100-Mbps switch prices is high.

The AT-8525 had the lab’s second-favorite Web interface and the lab’s
least-favorite console software. Just configuring the IP address was an exercise in
ingenuity and patience. Once everything was set to enable Web access, however, I was
pleasantly surprised by the management tools in the built-in Web interface.

Networking powerhouse Cisco’s Catalyst 2924M Series XL turned in the lowest scores
on throughput and latency tests although its configuration was as identical to the other
tested systems’ as the lab could make it. Time constraints did not permit me to meet
with a Cisco engineer to resolve the issues. The lab will likely take a second look to
find a fix or an explanation.

Performance, luckily for Cisco, was only part of the evaluation. It delivered the goods
in other areas such as connectivity options, management and scalability. Many of
Cisco’s technologies are proprietary, however, which causes problems in heterogeneous
environments.

Management, whether through Cisco’s own software, the Web or a Telnet session, was
easy across the board. The Web interface was eye-pleasing but had trouble submitting
system changes. Clicking the apply button usually worked, but about 30 percent of the time
it failed to record the change. Otherwise, the Catalyst 2924M could easily fit into a
management framework where other Cisco devices are in place.

Overall, the Cisco switch met or exceeded the lab’s expectations in every area
except performance.

The Compaq Dual Speed Switch SW3322, the most straightforward of the bunch, pretty much
lacked bells and whistles. An excess of extras often hurts a product, but having very few
also does. The SW3322 showed average performance, average management features, and little
in the way of scalability or upgradeability options.

That would be fine if it also had the lowest price, but the 3Com and HP switches were
cheaper and had many more features. The NetVantage switch on which the Compaq SW3322 is
based has two uplink ports, unlike Compaq’s version.

The SW3322 was notable for the most dedicated buffer memory per port, which helped
surprisingly little. The switch did support all common Layer 2 standards. Web management
mostly was acceptable, but I preferred the greater depth of bundled Compaq Networking
Management Software.

Overall, any of the five switches would be an improvement over shared-hub networks.
Whether you need switching for high-end desktop PC use, wiring closet hub concentrations
or server farms, you can expect performance and prices to improve in the near future.

That makes the Reviewer’s Choice remarkable. The HP ProCurve 2424M seems to be
ahead of the curve in every respect. With a per-port price around $54 and solid
performance and management, the 2424M clearly is suited for more uses than any other
switch the lab tested. 

The GCN Lab collected millions of data points for this review, one of the most
technical and thorough the lab has ever done.

I evaluated the performance, management consoles, vendor-specific software and embedded
Web servers of five 10/100-Mbps autosensing Ethernet workgroup switches. I looked at their
model options and configurations and analyzed how much of an upgrade path each switch
could offer.

The complete raw results appear on GCN’s Web site at www.gcn.com. Keep reading to
see how the lab came up with the GCN PowerIndex for switches. Because of the large volume
of results, I include here only a select group that illustrates the greatest differences
among the switches.

The lab invited 12 companies to participate in the review. Five companies sent
products.

The lab measured performance using the SmartBits 2000 Performance Analysis System from
Netcom Systems of Chatsworth, Calif., and a SmartBits 10 slave chassis. The lab’s
test methods are based on the Network Working Group’s Request for Comments 1242, RFC
1944 and RFC 2285 specifications. See Netcom Systems’ site at www.netcomsystems.com
for more information.

A switch’s performance, as determined by Netcom’s SmartApplications and
Advanced Switch Tests (AST) suite for Ethernet, accounted for 30 percent of the GCN
PowerIndex score.

SmartApplications examined the throughput, latency, packet loss and bursty traffic
performance. The AST suite applied X-Stream, head-of-line blocking, fan-out switching and
illegal frame handling tests.

Manageability accounted for 20 percent of the GCN PowerIndex score. Simple Network
Management Protocol compliance and Telnet access were basic requirements. Switches
received extra points for vendor-provided management software or management by standard
Web browsers. The lab also considered other manageability issues such as device stability.

For vendors that had proprietary management software, I evaluated how many platforms it
could manage. I looked for Spanning Tree Protocol support, Remote Monitoring groups, types
of IP configuration available, Trivial File Transfer Protocol, and out-of-band and in-band
management.

Vendors that relied on proprietary management protocols or technologies received extra
points for providing the capability, but they lost points if they failed to support an
existing open standard.

A third set of scores depended on scalability options, upgrade path specifications and
reliability. The lab considered each switch’s ability to grow beyond its base
configuration and accept asynchronous transfer mode or Gigabit Ethernet port uplink
options. Stacking with other switches without using up any of a unit’s 24 10/100-Mbps
ports also counted. The existence of port aggregation or trunking features figured in the
score, as did the number of possible aggregated pairs.

Switches that complied with the Institute of Electrical and Electronics Engineers’
802.1Q virtual LAN standard got extra points. The number of media access control addresses
a switch could handle also counted, as did Layer 3 switching functions.

Price and value determined 30 percent of the GCN PowerIndex score. The lab considered
chassis price, price per port and price of options. Features such as good documentation,
ease of configuration and other valuable capabilities raised the score.

Netcom’s SmartBits 2000 unit and SmartBits 10 port extension chassis gave the lab
a total of 40 ports for testing, 20 ports per chassis. The company’s SmartWindows,
SmartApplications and Advanced Switch Tests for Ethernet applications were crucial to the
lab’s review.

SmartWindows configures and modifies SmartBits’ modular cards that provide
connectivity to test devices. SmartApplications automates testing under the RFC 1242
Benchmarking Terminology for Network Interconnection Devices specification. I used it to
test switches; it also works for bridges and routers.

SmartApplications measures throughput by how many packets a switch can handle without
dropping any. It measures latency by how long a switch takes to forward a packet under
varying load conditions. It measures packet loss by how many packets are not forwarded
when the switch resources are low. SmartApplications’ back-to-back test measures the
capacity of the switch’s buffer.

The Netcom AST suite pushes a switch to its limits by sending dense, bursty traffic in
multiple directions. Although such a condition probably would never arise on a production
network, network engineers know it is wise to plan for the worst-case scenario.

The AST X-Stream test forces all switch ports to send and receive data at the same time
with occasional bursts. This demonstrates how well the switch handles bursty traffic,
memory buffering, and a variety of sizes and loads. The head-of-line blocking test shows
whether a switch has congestion control.

The fan-out test makes a group of ports send to another group. I applied it to eight
ports and to 20 ports to see how well the ports could send without loss while being
accessed by many other ports—an important question in designing a server farm.

Finally, the illegal frames test tells whether a switch incorrectly passes along any
undersized frames, bad checksums or other frames not up to specifications.