and Gigabit Ethernet have both evolved into mature and
robust LAN technologies. Each has unique advantages
that make it suitable for particular industries. Have
you made the right choice for yours? by Graeme K. Le
some respects, the competition between ATM and Gigabit
Ethernet (GbE) has more than a passing resemblance to
the battle between VHS and Beta technology. In those
days, both VHS and Beta video cassette recorders were
relatively expensive, of limited practical use, and
both newcomers were aiming to replace an older technology
(U-Matic format video tapes) which worked perfectly
well for most of the market.
In the end, VHS won the consumer market because users
were comfortable with a tape that looked like a miniature
version of what they knew (U-Matic cassettes), and that
VHS could accommodate four hours of recording on a single
tape. Beta won the professional market because of its
perceived superior picture quality and more functional
In the enterprise network marketplace today, ATM looks
like playing the role of Beta and GbE that of VHS. Note
that throughout this discussion we will use the term
GbE as an umbrella term covering both 1Gb and 10GbE
Telco's only option
For a telco, ATM is the only option for building very
largenational or globalnetwork infrastructures
that will carry a mixture of voice, data, and video
services with constant Quality of Service (QoS) and
predictable latency. On the other hand, non-telcos who
look for a high-capacity network infrastructure, see
GbE as just another form of the Ethernet they are using
Furthermore, the proponents of GbE point out that voice
and video can also be transmitted over Ethernet. Just
to complicate matters, both ATM and GbE seem to be evolving
into more mature and robust products. So how do you
choose between the two technologies?
Things to consider
The first thing to do is to figure out whether or not
you need either of them. In most cases ATM is not an
economically viable desktop technology and neither is
GbE. Even if you have a few users with a need for extraordinary
amounts of computing power, and who routinely handle
enormous data sets, it is likely that you will find
it more cost effective to install something like the
newer pocket supercomputers.
Such pocket supercomputers can have either a GbE or
an ATM interface, and you can have them in your data
centre along with a SAN, and then give specific users
high-end thin clients to access this machine via full-duplex
100BaseTx Ethernet links. CPU power is cheaper in large
chunks so there is no reason to be hauling huge amounts
of data across a network just to process it on a more
expensive platform. Especially when doing so would also
increase the cost of your network by forcing you to
accommodate Gb data streams to users' desktops.
In general, an average user with a thin client on their
desktop and a VoIP telephone handset will get by quite
happily with a full duplex 10Base-T link to their desktop.
Most PCs for business users have 10/100BaseTx Ethernet
NICs built-in, so add a VoIP phone and you will still
find that, at worst, full duplex 100BaseTx will do the
job. The key in both cases is to build a backbone system
that will support non-blocking full duplex connections
and guarantee both stable latency and enough bandwidth
for both the PC and the VoIP 'phone line'.
As far as desktop video conferencing is concerned, it
basically places the same constraints on the network
as VoIP, but requires more bandwidth. Once again 100BaseT
full duplex Ethernet is usually more than sufficient.
Given that you can get by comfortably with 10/100BaseTx
Ethernet to your users' desktops, the next question
is what will you need as a backbone technology for the
links between the local hubs and the hubs in your IT
glass house? From this point there are several choices
you can make using ATM, GbE or neither of them. Most
of these choices center around how much cable, specifically
fiber optic cable, you want to pull.
How much cable to pull?
A basic design rule for any network is that you use
copper UTP to the desktop and fiber between your hubs.
Short haul, multi-mode fiber is relatively inexpensive;
single mode fiber of the sort used in campus applications
is more expensive, but not unreasonably so.
Costs change drastically at the ports to which the fiber
is attached. Fiber versions of 100 Mbps Ethernet are
considerably cheaper than either GbE or ATM ports. Thus,
you basically have to decide on whether you want to
install lower cost 100 Mbps fiber ports, or less costly
but faster GbE or ATM ports.
Using 'lesser' 100 Mbps fiber Ethernet means segmenting
your network. The incidental effect of this segmentation
is arguably greater reliability, but there is a limit.
If your average user's desktop has a simple 10BaseT
link to it, effectively you have a flat Ethernet, and
you typically have 20 users per 10 Mbps segment. In
this case a single 100 Mbps full duplex link can support
up to 200 users without over subscription.
The reasoning for this is that each 10BaseT segment,
which is half duplex, can only present 10 Mbps to the
10/100 switch controlling the full duplex backbone link.
Each segment can be either transmitting or receiving
at any one time and hence the 100 Mbps full duplex link
can support up to 10 segments, or 200 users.
Obviously if you run this sort of network, the use of
GbE or ATM is probably pointless, but what if you need
to guarantee 10Base-T full duplex to each desktop? In
this configuration a 100 Mbps full duplex link will
support just 10 users without oversubscribing. In most
cases you will oversubscribe, depending upon your traffic
loads, by about 5-10 to one. In other words, about 50
to 100 users. Usually the number of users will be about
60; where VoIP is running, 40 users is probably as much
as you want to try.
Add desktop video conferencing to the mix and oversubscribing
becomes a very bad idea. Obviously, as the amount of
users you can support per link falls, you are faced
with the choice of upgrading your existing fibers to
GbE or pulling more fiber.
In the average TCP/IP network, ATM with either OC-3
(155 Mbps) or OC-12 (622 Mbps) will not be an economical
option, since the links are more expensive than GbE,
as are the hubs and their switch engines.
Take into account the fact that you will have to pay
a premium for your GbE switches, as you will need support
for 802.1p, 802.1q, etc., to carry your VoIP and video
conferencing traffic with the necessary QoS. Naturally,
ATM gives you the QoS support you need automatically.
By now, you are probably thinking that only telcos would
use ATM, but this is not necessarily so. The examples
we have discussed implicitly assume, at worst, a fairly
even mix of data, VoIP and video conferencing. In most
company networks, the bulk of the traffic is simple
data which can cope with reasonably wide variations
in latency. Even if you add VoIP, it is likely to constitute
around a third of your traffic, albeit with much tighter
constraints on variance in latency.
As for desktop video conferencing, it is just not something
a lot of people do. In most cases, video conferencing
is done between specifically configured conference rooms
at different sites, and it is often done via ISDN rather
than the company's TCP/IP network.
Video heavy traffic
But what if you had a network where video streams were
the bulk of the traffic, voice was the major secondary
application and data was only a relatively minor concern?
For example, consider a video production facility with
several studios, each with several digital video cameras,
a shared post production facility, shared special effects
shop which specialize in computer graphics, and an administration
center with a central PABX, all grouped on one campus,
or several in one city.
A thing to note is that a digital video camera produces
a data stream which occupies a lot less bandwidth than
a modern office PC can generate, and an audio stream
occupies even less bandwidth. Video cameras produce
streams of data, unlike the bursty traffic which a PC
generates. And there are going to be a lot less users
in this scenario than there are PCs in an average office
A typical TV studio has less than a dozen cameras, all
of which will not be in use at once. One floor of an
office block can have hundreds or more PCs, all of which
will be in use at once. We could use GbE or ATM in the
studio situation, but ATM is likely to be a more natural
choice, particularly when one may have to send a video
stream to a remote site via a telco's land line service.
Chances are that the telco can sell the studio a wide
area ATM service which can be directly interfaced with
the studio's ATM network.
It is probably true that most networks are going to
see GbE with its various enhancements as their most
cost effective and straightforward choice. But in cases
where QoS, streaming support, and the ability to directly
interface with a carrier's wide area services are more
of a priority than simple bandwidth, ATM is still arguably
the best choice.
Graeme K. Le Roux is the director of Moresdawn (Australia),
a company which specialises in network design and consultancy.
He can be reached at firstname.lastname@example.org