CompactPCI is a remarkable success story. In
1997, it was a $5 million industry focused on industrial control. Today,
it is a $400 million to $500 million market, growing 100 percent each
year, mostly the result of telecommunications applications. It's been a
case of the right technology in the right place at the right time. In
recent years, the telecommunications industry has come under increasing
competitive pressure. At the same time, data networking concepts threaten
traditional circuit switching, and computer industry business practices
threaten 120 years of telecom tradition. Today, telecom equipment vendors
can't afford to vertically integrate, designing every component that goes
into every solution. It takes too long and misses the opportunity to
leverage other industry innovators. As a result, all of the major telecom
equipment vendors are struggling to find ways to utilize standard
components and mass-market computer technology while still delivering
highly reliable telecommunications solutions.
ENTER CompactPCI
CompactPCI began as a robust mechanical and electrical chassis that
was compatible with mass-market PC processors and, more important, with PC
software. As the result of work done in 1997 and early 1998, CompactPCI
added a standard for hot swap and a telecom-specific configuration using
the ECTF's CT Bus (H.110). Hot swap provides the ability to remove boards
from or add boards to a running system, which substantially improves
system availability. The CT Bus provides a telecom bus, with redundant
clocks, that also supports hot swap.
Suddenly, there was a way for telecom equipment vendors to leverage
standard industry components and mass-market computer software, but to do
so in a telco-grade environment. CompactPCI immediately became important
for software-intensive systems such as new application servers and media
servers. But with a base established, CompactPCI has been deployed in a
much wider range of telecom applications.
At this point, CompactPCI has displaced VME for new design wins in
telecom. More important, it has displaced a great many proprietary chassis
in systems from the major telecom equipment vendors. But, even though it
is a booming business and growing at a great rate, the CompactPCI market
could be bigger. There are some applications that need more capacity,
either PCI or CT Bus bandwidth. And, increasingly, there are applications
that need packet backplanes, either in addition to PCI or instead of PCI.
Today, people solve these needs by utilizing parts of the CompactPCI
environment and then adding vendor-specific extensions. Could we have a
few common agreements that help a significant fraction of the market?
Let's look at three areas where common extensions might be useful: PCI
Bus, CT Bus (H.110), and packet backplanes.
PCI AND PCI EVOLUTION
What's In A Name?
The widespread adoption of the PCI bus formed the basis for
CompactPCI and the name of the bus has been embedded in name of the
industry. But change is always on the horizon. Two years ago, Intel
announced a new I/O scheme, NGIO. And a competing consortium of major
computer vendors announced FutureIO. After moves towards a compromise,
started last summer, a single new I/O scheme is now being developed under
the name Infiniband. While initially targeted at high-capacity peripherals
and box-to-box connections, Infiniband is slated to replace PCI
eventually.
The key feature of CompactPCI has been its ability to leverage
mass-market computer software. If mainstream PCs do move to Infiniband,
then CompactPCI must follow. But does the name CompactPCI need to change?
Perhaps not. Even today, not everyone using "CompactPCI"
components is using PCI. What is needed is the clear message that
CompactPCI is much more than PCI.
Infiniband
Infiniband, like many other high-speed I/O interconnects, is based on
high-speed serial links connected together with switches. Adding switches
allows more links, which means more capacity for the I/O fabric. At least
in theory, the I/O fabric can scale indefinitely.
Unfortunately, as it is currently envisioned, Infiniband is not
compatible with the existing PCI bus and therefore requires new I/O device
driver software and new I/O management software in the operating system.
So, adoption will take much longer than today's hype would have us
believe.
Consider the evolution of PCI, for example. The PCI standard was
completed in 1992, PCs with some PCI slots were widespread by 1996, and
PCI-only computers were dominant by 1999. And PCI didn't require all new
I/O software. The Universal Serial Bus (USB) provides an alternative
example. It was defined in 1994, but it had to wait for Windows 98 before
it was even useful. Today, it's included on all new computers, but it is
only just beginning to be used to connect some PC peripherals.
We can expect a comparable adoption period for Infiniband. Assuming the
initial specification is delivered later this year, we can expect to see
built-in operating system support in 2004 and ubiquitous adoption in 2008
or 2010. The clear message should be that CompactPCI leverages mass-market
computer technology and therefore will track the computer industry as the
industry evolves. But what should the CompactPCI market do in the interim?
Scaling PCI In The Interim
Some applications need to boost capacity on today's PCI bus and/or
carry the bus between shelves in a multi-shelf CompactPCI system. In a
single chassis, some measure of relief is obtained by localizing traffic
behind PCI-to-PCI bridge chips (from Intel or PLX Technology, for
example). Beyond this approach, several schemes have emerged that promise
to bridge multiple PCI segments with more capacity over longer distances.
PLX acquired high-speed ring technology, by their acquisition of
Sebring Networks, and have announced a Gigabridge -- a bridge from PCI to
their high-speed ring. And StarGen recently announced PCI bridging over
their high-speed serial switch fabric. This is a switch fabric much like
Infiniband plans to be, but PCI compatible. A potential advantage of the
StarGen solution is that, like Infiniband, it scales indefinitely while,
without extra complexity, the ring-based solution can only scale to the
capacity of the ring.
GREATER CAPACITY CT BUS (H.110)
While some are looking for more capacity on the PCI bus, many others
want more capacity on the telecom bus within the CompactPCI chassis. The
H.110 bus was the result of joint efforts by the ECTF's Hardware Working
Group and the Telecom Interest Sub-Committee (TISC) of PICMG. The
CompactPCI version of CT Bus (H.110) adds some key enhancements over the
PC version (H.100), notably the elimination of ribbon cables for
connecting boards and the addition of redundant clocks and support for hot
swap.
The TISC specification does its job extremely well as long as its 4,096
timeslots are adequate. Unfortunately, there is no roadmap beyond the
4,096 timeslots of today's CT Bus. This limits many people in deploying
CompactPCI solutions for high-end applications. But, even without a
standard, custom solutions are emerging.
One solution is a backplane that segments the H.110 bus into smaller
sections with multiple segments connected to a single high-capacity,
proprietary line card. Another approach creates multiple secondary H.110
buses to connect small groups of boards, while reserving the primary H.110
bus on the backplane for interconnecting these board groups. Other
solutions involve using the H.110 bus for a subset of the application and
moving the rest of the traffic onto a proprietary packet data bus or other
paths. Clearly, a common solution could leverage CompactPCI into new
markets.
The most interesting new entrant is the joint announcement by StarGen
and Lucent Microelectronics of their plan for a bridge from H.110 to the
StarGen switch fabric. This will provide a scalable way to interconnect
individual CT Bus segments either within one chassis or between chassis in
a system.
THE NEED FOR A PACKET BACKPLANE
Another area that's wide open is packet backplanes. Currently there is no
generally agreed upon packet backplane for use within a CompactPCI
chassis. There are many proprietary solutions and various ideas being
floated. Solutions range for a few hundred megabits of bandwidth to 80 or
100 Gbps of packet transfer capacity. Many schemes put 100Base-T or
Gigabit Ethernet links on the CompactPCI backplane. A typical
configuration is two point-to-point links from each slot, one to each of
two (redundant) Ethernet switch cards. A more costly but higher capacity
and completely distributed approach uses a full-mesh of point-to-point
Ethernet links between every slot and every other slot with a full
Ethernet switch on every card.
Among other solutions, TranSwitch and AMP have application notes that
describe how to put the TranSwitch ATM CellBus onto the P3 connector of a
CompactPCI backplane. This provides roughly 1.2 Gbps throughput for ATM
cells or packets that have been mapped to ATM cells. There is an even
wider variety of custom backplanes that leverage proprietary switch
fabrics. Finally, there are proposals based on emerging switch fabrics
such as RapidIO, StarGen, and Infiniband.
What, then, is needed to take the CompactPCI market to the next level?
One, or a few, standard packet backplane configurations endorsed by PICMG.
They should follow the CompactPCI philosophy of leveraging mass-market
technology and especially mass-market software. That is, a CompactPCI
packet backplane should work with existing software to the maximum extent
possible. A single, agreed upon Ethernet backplane is an obvious need. A
highly scalable PCI-to-PCI bridging technology, like the StarGen fabric,
could be an interesting packet backplane. And eventually, once there is
software support and widespread deployment, Infiniband might also serve as
a packet backplane for CompactPCI.
A PATH FORWARD
Today, CompactPCI has found a sweet spot in the telecom market in
systems with a few dozen and up to a thousand or more traditional voice
circuits. However, for systems with more than 2,000 channels, and for any
packet-based system, today's CompactPCI requires extensions -- extensions
which so far have been vendor specific. It's still good business. The
CompactPCI market is already much more than just PCI computers for
telecom. But the market could be much, much larger if we had industry
agreement on the more common extensions.
What we need today is a clear roadmap for several new optional
backplane configurations that address packet switching, added
circuit-switching capacity and extended PCI bridging. The good news is
that there are a lot of people in the CompactPCI community thinking about
these issues.
We need agreement on an Ethernet backplane as soon as possible. Beyond
that, my personal bet is on a backplane using the StarGen switch fabric.
The StarGen switch fabric has the advantage of bridging both the PCI bus
and the H.110 CT Bus. As a result, it's a high-speed switch fabric that
can extend existing telecom backplanes and thus work with all of the
existing CompactPCI telecom boards. Even more important, it leverages the
existing base of software for PCI computers and H.100 telecom
applications.
Even with no further action by PICMG, the CompactPCI market will boom.
But with just a few additional, agreed-upon specifications, the market can
grow much more rapidly.
Brough Turner is senior vice president of technology at Natural
MicroSystems. For more information, call Natural MicroSystems at
508-620-9300. E-mail to the author (addressed to [email protected])
is also welcome.
[ Return
To The October 2000 Table Of Contents ] |