Carriers and high-end service providers have always demanded up to "six nines" (99.9999%) reliability from their switching and computing systems. That's why the mass-produced PC has taken so long to become a player in the stringent operating environments found at the top of the convergence chain.
In early 1994, "RuggedPCI" by Ziatech (now part of Intel) was conceived of as a way to bring relatively inexpensive, open-architecture, highly efficient, reliable, and easily upgradeable equipment to circuit- and packet-switched networks. Swiftly renamed CompactPCI, the new bus gave cost, effective PCI circuitry a rugged home on Eurostyle cards and connectors. Under the standards brought forth from PICMG (PCI Industrial Computer Manufacturers Group), hot-swappable boards appeared to satisfy the quick maintenance requirements of telcos.
Convergence-friendly options for cPCI include -48 volt power supplies to make cPCI compatible with other central office equipment - most of the -48V cPCI technology suppliers offer dual or even triple N+1 redundant power supply configurations. Also adopted into the cPCI spec was the ECTF's (Enterprise Computer Telephony Forum) H.110 CT bus for computer telephony resource and I/O boards. H.110 provides an intra-chassis time-division multiplexed digital telephony highway with 4,096 timeslots.
It's taken years for all of the specifications to be approved. Now that cPCI is nearing maturity, however, it's becoming apparent that the technology still suffers from certain limitations.
cPCI PROBLEMS
Before the CompactPCI craze swept through telecom, medical, industrial and military sectors, the Versa Module Europa (VME) served as the main computing platform for high-end telecom and other real-time computing systems. VME has been around since 1981, supports a huge number of protocols, and has a tremendous installed base. It's an open-ended, flexible computer backplane bus with a 32-bit wide data path that, like CompactPCI, is built upon the rugged Eurocard standard.
But unlike cPCI, VME has a pretty efficient interrupt scheme that allows you to put a full 21 slots on a backplane - it was only this year that APW Electronic Solutions introduced their StealthBridge backplane that supports 21 cPCI cards and a full complement of rear-mounted I/O transition cards - a gain of two or more slots per system. This achievement becomes more impressive when you realize that two extra slots translates into eight more T1 lines or 256 full duplex phone lines.
Also, in a large cPCI system with multiple CPU cards, one slot - the System Slot - must provide functions such as clock generation, interrupt handling, and bus arbitration, which means the System Slot can be a single point of failure. And drivers compatible with the operating system working with the Slot One controller must also exist in each chassis unit. This means that you need OS-specific drivers, which can complicate product development. On the hardware side, possible solutions include moving these functions to another plug-in card or rear panel transition module, and making them redundant. Further work will also be required to standardize and adapt drivers across multiple OS platforms.
Then there's the fact that standard cPCI subsystems communicate via a single cPCI bus interface. This is a single point of failure, so the entire system becomes inoperable if the interface controller on one subsystem siezes the bus. Also, if you bend a pin while hot-swapping a card, the board can short out and the whole system can go down. The PCI bus itself is a point of failure.
More expensive systems can achieve fault tolerance by having redundant cPCI buses, but the chassis' width limits the number of slots, as does the fact that the PCI bus (a shared bus architecture) has a theoretical throughput limitation of 533 MBps over only five slots on a 66 MHz bus. This is far less bandwidth than, for example, the latest forms of Ethernet offer.
Moreover, the H.110 bus - with just 4,096 timeslots - has begun to look pretty anemic. A big, present-day system could use ten times as many timeslots, just for starters. More sophisticated distributed processing and box-to-box communications schemes are clearly required.
R&D TO THE RESCUE
These limitations have spurred a tremendous R&D effort that is transforming CompactPCI equipment. Soon, CompactPCI chassis and boards may look physically the same as ever, but the data traffic moving along the signal traces on the backplane will no longer adhere to PCI signaling characteristics.
In short, it's CompactPCI without the PCI. Of course, back-compatibility with cPCI H.110 boards will be preserved for some time, so legacy installations won't be in a bind.
These new technologies take on various forms (see the sidebar on Backplane Politics for a look at the PICMG 2.13, 2.16, and 2.17 specifications).
Most notably, as you review the detailed product roundup section, you'll see mention of the so-called "packet switched backplanes," which is the first of these new technologies to appear in a CompactPCI form factor. Packet-switched backplanes are based on cost-effective, well-known networking technology. Since Ethernet now has a greater bandwidth than the PCI bus, many vendors decided to build a "LAN in a box" and connect the boards with it, following PICMG's 2.16 specification.
Two years from now the PCI bus will give way to advanced "serial and pseudo-serial switching fabrics" such as InfiniBand, currently under development by companies such as Compaq (recently acquired by HP), Dell, HP, IBM, Intel, Microsoft, and Sun Microsystems. As the alleged "heir apparent" of future computer architectures, InfiniBand will give the server world a much-needed boost in terms of bandwidth capability and fault tolerance. Unlike PCI, InfiniBand's links can come out of the box, enabling flexible network connections that can scale easily and yet provide fault tolerance.
INFINIBAND
We spoke with Kevin Deierling, vice president of product marketing for Mellanox Technologies Inc. (Santa Clara, CA - 408-970-3400, www.mellanox.com), a company that's taken the first steps toward linking the InfiniBand Future with the PCI Past. Mellanox' InfiniBridge technology was released in January of this year.
A board with an InfiniBridge chip looks like a standard PCI card (a cPCI version is in development), and indeed it interfaces to the circuitry of the legacy PCI bus. But the InfiniBridge can support 10 Gbps links traveling over copper cables up to 17 meters (55.7 feet) in length. The 10 Gbps link is sent as four 2.5 Gbps pipes using what's called "byte-striping" - the first data byte is sent on the first twisted wire pair and the next byte on the next pair, and so on. It's reminiscent of time division multiplexing, except that the bytes are stripped across the channel in parallel instead of in sequence.
