Networking for the New Millennium: After ATM, What?
telecoms-mag.com
Cruising the Information Superhighway
Jim Mollenauer
In retrospect, it was obvious all along: Small computers are
faster than large ones. Large computers send signals longer
distances, and the wiring has larger capacitances to charge
up. So microprocessors are inherently faster than
mainframes. It just requires the ability to put a million or more
components on a single chip. So when you open up a
mainframe, deep in the heart of a large cabinet you will
usually find a microprocessor.
Can we apply this to network technology? Can we derive any
rules from these very general principles? Yes, at least one:
The less often new switching decisions are made by
interpreting a data unit header, the faster the network can run,
or the less complex the equipment required. From this rule, it
appears that a new technology known as dynamic
synchronous transfer mode (DTM) may offer better
optimization for many applications than anything we have
seen before.
Consider the need for flexibility in network architectures.
Demands change, and the network needs to reflect this. For
example, systems like SONET process each byte differently.
Here multiplexing is on a byte basis, and the receiver may
send each received byte to a different destination. The time
for a decision is very limited, and the capabilities are
restricted to repetitive demultiplexing, with all channels
restricted to constant bandwidth.
ATM permits changes once per cell: After every 48 bytes of
payload, the switch is able to do something different. This
enables more complex capabilities, but the complexity has
been hard to manage and exacerbated by the need to
coexist with other technologies.
IP does better than ATM in this area because it makes its
routing decisions once per packet rather than once per cell.
In the early days, the one group noticeably lacking in
enthusiasm for ATM was supercomputer users. They
recognized that the larger the data unit, the faster it is, and
that the ATM cell was too small. Today, with desktop
computers running at speeds rivaling the supercomputers of
a decade ago, these considerations apply widely.
The optimum situation, however, is one where the network
paths change only as often as the applications change. This
is the principle of DTM. DTM operates by time division
multiplexing small (64-bit) slots on a shared ring or bus in a
way that permits very fast reassignment of bandwidth when
needed. Assignments are good until changed, unlike ATM or
packet protocols, where the internal functions of the switch
are, in effect, reset after every cell or packet. The result is that
the control process must intervene only when the applications
require it; in the meantime, the destination equipment picks
out the correct data based on simple slot-counting
procedures.
Counting is very inexpensive to implement, compared to
alternatives like ATM. There, it is similar to SONET
demultiplexing, but the granularity is much finer. A single slot
per 125-microsecond frame represents 512 kbps, a hundred
times smaller than the minimum OC-1 at 51 Mbps. This fine
grain makes it ideal for multiplexing disparate traffic over a
metropolitan or regional backbone. Examples of this include
mixed digital video and cable modem data traffic for the
cable TV industry, and interconnecting the base stations in
wireless networks for combined voice and data.
The fast response of DTM is achieved by allowing each node
to keep an inventory of channels (sets of time slots) going to
a variety of destinations. If a request comes in that can be
accommodated by slots that are already on hand, allocation
is immediate, and the response is as fast as any router or
LAN switch can provide. Buffering is not needed at any
intermediate node: Data moves in and out at the same
speed for any one connection.
If there are not enough slots on hand for the needed
destination, the node canvasses its neighbors until it collects
enough to set up a channel with the requested rate.
With this mechanism, control operations are needed only as
frequently as applications change: perhaps once for a 1-Mb
file transfer, rather than a thousand times for typical IP
packets or 20,000 times for ATM cells. For video, once an
hour or two might suffice for a constant-rate MPEG movie or
for an overall MPEG stream containing statistically
multiplexed programs.
DTM can work especially smoothly with IP, since IP routing
protocols such as Open Shortest Path First (OSPF) can be
used to find the routes in a set of interconnected DTM links.
DTM did not appear overnight. It was nurtured in the face of
ATM's popularity and the more recent view that IP will take
over the world. Credit for developing it should go to Ericsson,
which has funded research on DTM for over 10 years at the
Royal Institute of Technology in Stockholm, and to two
Swedish start-ups spun off from that effort. Real products are
starting to roll out from Dynarc (now located also in Silicon
Valley) and Net Insight. DTM standardization is getting
underway at the European Telecommunication Standards
Institute (ETSI), paving the way for acceptance by network
operators. Just as LANs were the big news of the ?80s and
ATM for the ?90s, DTM may achieve that status in the first
decade of the new millennium.
Jim Mollenauer (jmollenauer@technicalstrategy.com) is a
contributing editor with Telecommunications© magazine
and president of Technical Strategy Associates in Newton,
Mass., providing consulting on high-speed networks,
including ATM, cable modems and wireless. |
