SI
SI
discoversearch

We've detected that you're using an ad content blocking browser plug-in or feature. Ads provide a critical source of revenue to the continued operation of Silicon Investor.  We ask that you disable ad blocking while on Silicon Investor in the best interests of our community.  If you are not using an ad blocker but are still receiving this message, make sure your browser's tracking protection is set to the 'standard' level.
Technology Stocks : Cabletron Systems (CS: NYSE) -- Ignore unavailable to you. Want to Upgrade?


To: BMcV who wrote (5267)11/10/1998 11:53:00 AM
From: Kenneth E. Phillipps  Read Replies (2) | Respond to of 8358
 
Excerpt from DataCom re Gigabit Ethernet Switches - Where was Cabletron?

"In fact, when we called for products last August, only four vendors could deliver all the features our readers
say they really need on the corporate backbone: Bay Networks Inc. (Santa Clara, Calif.), Extreme, Foundry,
and Packet Engines (see Table 1"

---Brainier Boxes -- The latest gigabit Ethernet backbone switches have more than speed to offer-like
intelligent routing and QOS
David Newman, Tadesse Giorgis, and Farhad Yavari-Issalou

Along with NSTL Inc. (Conshohocken, Pa.), Data Comm evaluated five gigabit Ethernet backbone boxes and
found that all have what it takes for backbone service: high port density, intelligent routing, quality-of-service
features-and blazing speed. We did a complete workup on each switch, looking at 15 different measures of
backbone suitability and offering far more traffic than we have in any past test. And the numbers were
staggering: Most of these switches routed IP traffic at up to 36 million packets per second-three times better
than what we offered in previous gigabit Ethernet tests. What's more, all five boxes proved adept at fielding
traffic from thousands of attached devices.

Results like that made it hard to pick a winner, but in the end we gave Tester's Choice awards to two
products. The Powerrail 5200 from Packet Engines Inc. (Spokane, Wash.) aced nearly every performance
routine, and it also offers a long and impressive features list. And the Bigiron 4000 from Foundry Networks
Inc. (Sunnyvale, Calif.) posted stellar numbers as well-boasting the highest address learning capacity among
the field. Both switches rank as the fastest we've ever tested.

Roster Rundown

A quick scan of our test lineup is all it takes to notice the absence of some major players. Such vendors as
Cisco Systems Inc. (San Jose, Calif.) and 3Com Corp. (Santa Clara, Calif.) have gigabit Ethernet boxes, but for
the most part these are Layer 2 devices, edge switches with one or two high-speed uplinks, or both. Cisco,
perhaps mindful of the revenues generated by its routers, doesn't have a gigabit Ethernet switch with both IP
routing and high port density. 3Com has the Superstack II Switch 9300, but it's strictly a Layer 2 device.

In fact, when we called for products last August, only four vendors could deliver all the features our readers
say they really need on the corporate backbone: Bay Networks Inc. (Santa Clara, Calif.), Extreme, Foundry,
and Packet Engines (see Table 1

Switches from these vendors are suited for even the largest enterprises, with support for up to 73 gigabit
Ethernet ports and dozens or hundreds of Ethernet and fast Ethernet ports. And Packet Engines offers
interfaces for ATM, FDDI, and packet-over-Sonet networks. What's more, all the devices we looked at route
IP. Switches from Foundry and Packet Engines also route IPX, while Bay says it will add IPX support to its
Accelar 1200 by year's end.

The list of added features doesn't end there. All vendors also offer some type of QOS mechanism-essential on
the backbone. Switches from Bay and Packet Engines, for instance, can prioritize traffic based on settings in
the IP header's TOS (type of service) field. The beta version of Extreme's Black Diamond 6800 prioritizes
flows based on TCP or UDP (user datagram protocol) port number, and Packet Engines says its switch can
prioritize based on the signature of a given application. All vendors also support the IEEE's 802.1p and
802.1Q specifications for defining VLANs (virtual LANs) and prioritizing among them.

Finally, all the switches support four groups of RMON (remote monitoring) on every port, allowing net
managers to get a view of MAC (media access control)-layer switch statistics and history and set alarms
without dedicating costly RMON probes. (But none support all groups of RMON or RMON 2.)

Speed's the Thing

Those functions all help make these switches right for the backbone, but speed is still what counts most. To
find out how they stand up to heavy traffic, we deployed Smartbits analyzers from Netcom Systems Inc.
(Chatsworth, Calif.) to offer wire-speed traffic on up to 24 gigabit Ethernet ports at a time. That's the highest
density yet in a Data Comm test. But we also offered lots of different traffic: In all, we conducted 15 different
tests of Layer 2 and 3 forwarding rates, broadcast latency, Layer 2 address learning, head-of-line blocking, and
routing lookup performance (see "Test Methodology"). We ran these tests with eight gigabit Ethernet ports,
then repeated them with 24 ports on the boxes from Extreme, Foundry, and Packet Engines.

Bay's Accelar 1200 was the odd box out, dropping nearly 47 percent of traffic when we offered a 100 percent
load. Packet loss fell somewhat with lighter loads, but even at 80 percent it dropped nearly 34 percent of
packets. And the Bay switch continued to drop traffic even when we offered 1,518-byte packets: It lost
nearly 34 percent of a wire-speed load and around 20 percent of an 80 percent load.

Packet Engines offers a function called on-demand learning that gets at the flooding issue and could allow its
switch to learn a higher aggregate number of addresses. Instead of flooding unknown addresses to all ports,
new addresses are cached by a management module and forwarded only to other ports within the same VLAN
(virtual LAN) or broadcast domain. While we verified that on-demand learning reduces the amount of flooded
traffic, it was beyond the scope of the test to find out whether the approach increases the aggregate number of
addresses that can be learned.

Over in the address learning rate tests, Foundry's Bigiron 4000 once again was the leader, learning more than
70,000 addresses per second (see Figure 3). The smaller Foundry switch was next, learning at a rate of nearly
50,000 addresses per second. The other three learned between 4,800 and 5,800 addresses per second. Extreme
says it's still working on this part of the beta software we tested; it actually expected to see a far lower rate.

Layered Takes

We also repeated the test with 24 ports-which meant devices would have to route at an aggregate rate of
nearly 36 million pps.

With eight ports, most products turned in perfect performances (see Figure 4). The Extreme and Packet
Engines boxes forwarded 100 percent of packets without loss, as did Foundry's Bigiron 4000. Foundry's
Turboiron/8 dropped 4 percent of packets.
The results also were impressive with 24 ports. In fact, the Extreme, Foundry, and Packet Engines switches
were perfect, routing nearly 36 million IP packets without dropping one. These numbers are extremely
encouraging when put in the context of current routing capabilities; even the largest conventional routers today
move traffic at aggregate rates well below 1 million pps. In other words, network architects who are looking to
strengthen their backbone networks need look no further than these devices.

Wide View

Our final Layer 3 test involved routing packets to any IP address, not just those on the corporate campus. Of
course, this is a job usually handled by the WAN router connecting the campus to an ISP (Internet service
provider), but Bay, Foundry, and Packet Engines market their boxes for use in ISP sites. And that made the
test quite relevant.

We used a modified version of the so-called Torrent test, created by router vendor Torrent Networking
Technologies Inc. (Silver Spring, Md.) and Netcom Systems. The original test offers more than 1 million
unique IP addresses to the switch, a figure that's far too high to be meaningful for corporate campuses. Users
and vendors told us that campus routers might handle hundreds or maybe thousands of arbitrary lookups at a
time, but no higher. We reduced the size of the test so that we'd offer just 4,000 unique IP addresses, and if a
switch could successfully forward these we'd increase the load to 12,000. For both loads, we offered all
addresses at wire speed to one switch port and sought to have them forwarded to four outbound ports. We
measured addresses forwarded, and considered each packet delivered to be an IP address successfully learned.

Switches from Bay, Foundry, and Packet Engines handled both of the loads with ease. They all learned all
addresses without packet loss-except for Bay's Accelar 1200, which dropped an insignificant 0.78 percent of
offered addresses.

It was a different story for Extreme's Black Diamond. The vendor's architecture routes packets to known
addresses at wire speed, but packets with unknown IP destination addresses get sent to a management
module. The software-based processing that results takes far longer than simply forwarding the packet, which
is done in silicon. As a result, the Black Diamond learned only 1.46 percent of IP addresses at wire speed, and
the switch didn't learn all addresses until we reduced the offered load to less than 1 percent. This shouldn't be
taken as a knock. Extreme says its products are intended as LAN switches, not ISP backbone devices. On
corporate campuses, IP addresses are already known-and thus arbitrary address lookups just aren't an issue.

-David Newman is senior technology editor for Data Comm. His e-mail address is dnewman@data.com.
Tadesse Giorgis is a senior test engineer at NSTL Inc. (Conshohocken, Pa.). His e-mail address is
tadesse@nstl.com. Farhad Yavari-Issalou is a project manager at NSTL. He can be reached via e-mail at
farhad@nstl.com.

---

Thanks

Data Comm gratefully acknowledges the support of Netcom Systems Inc. (Chatsworth, Calif.), which
supplied its Smarbits analyzers with 24 of its gigabit Ethernet line cards and 32 multilayer 10/100 line cards;
AMP Inc. (Harrisburg, Pa.), which supplied multimode fiber cabling; and Extreme Networks Inc. (Cupertino,
Calif.), which supplied a Summit4 switch for prototyping of this test.

---

Sidebar-

TOP PERFORMERS

-Foundry Networks: High performance and low prices are the hallmarks of the chassis-based Bigiron 4000,
which learned more MAC addresses than any other product tested, aced the Layer 2 and 3 stress tests, and is
far less expensive than the other 24-port configurations we tested.

-Packet Engines: The Powerrail 5200 delivered letter-perfect results in our Layer 2 and 3 forwarding-rate
tests. It also boasts the most gigabit Ethernet ports per chassis, IP and IPX routing, and extensive QOS
capabilities.

---

Sidebar-

LAB TEST VENDOR PARTICIPANTS

Bay Networks

The Accelar 1200 has excellent Layer 2 and Layer 3 address learning capabilities, first-rate VLAN support,
extensive management features-and a performance problem when traffic rates rise to extreme levels. In our
stress tests, the Accelar dropped up to 50 percent of Layer 2 and 3 packets. Bay says packet loss isn't a
problem in production networks, where traffic rates are much lower over time.

Extreme Networks

Extreme's already impressive lineup of reasonably priced, unreasonably fast switches now has a chassis-based
entry with the Black Diamond 6800, which boasts up to 48 gigabit Ethernet interfaces. While the beta version
aced all our Layer 2 and 3 stress tests, there were a few rough edges: The switch learned only 4,700 MAC
addresses, far short of the record 131,000 addresses learned by an Extreme switch in an earlier test, and it did
so at a relatively low rate. The vendor was still at work on the switch's beta code at the time we tested and
says the production release should turn in higher performance numbers.

Foundry Networks

Both the chassis-based Bigiron 4000 and fixed-configuration Turboiron/8 put up impressive numbers. The
Bigiron is not only the least expensive of the 24-port gigabit Ethernet switches tested, it also made perfect
scores in all our Layer 2 and 3 stress tests and learned more MAC addresses than any other product we
looked at. The Turboiron/8 was nearly as good, showing only minor packet loss in one of our Layer 3 tests
but posting perfect results everywhere else.

Packet Engines

The chassis-based Powerrail 5200 delivered terrific overall results: It aced all of our Layer 2 and 3 forwarding
rate tests, moving up to 36 million packets per second without dropping a single one. It's also the only
product tested that supports ATM, FDDI, and packet-over-Sonet interfaces in addition to gigabit
Ethernet-and it also offers multiprotocol routing and several QOS mechanisms. It also can control the number
of MAC addresses sent to all switch ports, an important consideration in devices with high port density.

------

Previous