Since data traffic first showed signs of eclipsing voice in the late 1980s, telecom service providers have considered assassinating time-division multiplexing in broad daylight or gently euthanizing the time-slotting technology when no one was looking. But, like the constantly reviving undead, TDM networks have a way of coming back again and again after being declared dead and buried.

Time slotting was conceived as a way to channelize 64-kbit/second voice traffic, and the efficiency of multiplexing such traffic has been one of the factors keeping circuit-switched voice traffic reliable and resilient in the face of network breakdowns. But TDM also has played a role in aggregation at the network core.

Digital leased-line circuits, the T1/T3 hierarchy in North America and E1/E3 equivalents in Europe, are based on TDM. So are the Sonet payloads used in the fiber-based hierarchy of dual rings used throughout metropolitan areas. Again and again, packet afficionados have advocated schemes for cutting TDM out of the network, only to learn that carriers aren't prepared to make a complete cutover to Internet Protocol (IP) packets.

Asynchronous transfer mode (ATM) was the first carrier-designed concept meant to bring together data packets and circuit-emulated voice, though the size of ATM cells-53 bytes-represented a compromise that satisfied no one. To be sure, core networks transitioned to ATM across the board, but these ATM networks used the hit-and-miss available-bit-rate service and ATM Adaptation Layer 5 designed specifically for data packets. In almost all circumstances, continuous-bit-rate and other services designed for voice were talked about endlessly, and never implemented.

Using ATM bought the carriers some time in the 1990s for conversion to IP, and represented a better switching infrastructure at the core than frame relay. But there is some degree of irony in the fact that an updated frame-relay switching method, committed information rate, brought the older frame-relay switches closer to coexistence with TDM than most of the ATM networks that still exist today.

ATM's demise was due largely to the dominance of Ethernet traffic at the desktop: If LAN users and broadband-access clients would not accept ATM in the last mile, its role in the core was doomed. Since then, developers have attempted to address low-latency traffic from two directions.

One concept sought to make a packet-based datagram service "look and feel" like a circuit through clever traffic shaping and quality-of-service (QoS) prioritization. At the core of large metro and interregional networks, router makers advocated Layer 3 IP routing along with a new "tag switching" concept of defining IP flows called Multiprotocol Label Switching, or MPLS. Meanwhile, the manufacturers of the largest Layer 3 Ethernet switches for the enterprise tried out various forms of Ethernet-based quality-of-service queuing so that time-sensitive traffic could be carried from enterprise-switched LANs to core MPLS routers.

Reluctant to switch
If the corporate world had cut straight over to 100 percent voice-over-IP (VoIP) use, this topology would have made sense for everyone. Many networks are being installed in metropolitan areas today, and they could represent the future carrier topology of choice. But carriers told equipment designers that customers were reluctant to give up T1 and T3 services, and some of them were adamant about not giving up circuit-switched voice.

Continuing limitations of VoIP in handling 911 calls, fax services and network-powered phone calls has also made corporate clients more adamant about retaining TDM. So members of the Metro Ethernet Forum began rethinking the problem, working backward from the central office to the customer site.

Architectures from the likes of Overture Networks Inc. and Ceterus Networks Inc. are based on bonding TDM channels into Ethernet transport services, thus providing the same types of encapsulation offered in modern Sonet services, such as generic framing procedure and virtual concatenation.

There are also schemes for bringing timing information directly to packet switching. In early June, Zarlink Semiconductor Inc. (Ottawa) introduced a patented technology the company calls Timing Over Packet, in which a hardware processing engine adds timing and synchronization information to Ethernet, IP, ATM and MPLS packets. The Zarlink processors encode and transmit a master clock over the packet network and recover the clock at client nodes. Since Zarlink, formerly Mitel Semiconductor, was responsible for developing controllers for the time-slotted H.110 bus used in older enterprise telephony networks, OEMs are giving the Zarlink proposal careful consideration.

The ultimate irony of this reaching back to TDM comes in the effort to define next-generation backplanes for communication equipment. The Advanced Telecommunication Computing Architecture, a project promoted by the PCI Industrial Computer Manufacturers Group and other trade organizations, originally was intended to define architectures for the largest routers and Layer 3 switches at the network core. With the addition of the mezzanine-card-based MicroTCA, the ATCA family has been extended to all kinds of devices functioning at the network edge, including access routers, IP PBX systems, DSL access multiplexers and Ethernet switches.

A year ago, PICMG began working on an adjunct for both ATCA and MicroTCA called Internal TDM, or iTDM. Todd Wynia, vice president of marketing at Artesyn Technologies Inc.'s communications products group, said that "as ATCA began to be defined, we realized we lost something from the days of the H.110 bus: support for TDM services. Of course, the dream was that this would not be necessary in the 21st century. The reality is that telecom equipment, in most cases extending from customer premises to the network core, is going to need to support TDM with an internal bus."

The contention between the two design methodologies resides in how precisely packet services must emulate or encapsulate time-sensitive services. The Ethernet flow-monitoring and bonding schools base architectures on the assumption that physical-layer networks have so much bandwidth, and packet traffic has made such progress in being multiplexed according to flow, that retaining true time-slotting or clock references is largely unnecessary.

In a famous exchange three years ago on a QoS panel, World Wide Packets Inc. founder Bernard Daines called QoS traffic monitoring largely unnecessary, since "when you're in an Ethernet environment, you can always throw more bandwidth at the problem." The current dilemma of traffic oversubscription at the edge of Ethernet networks, which has spawned specialty semiconductor companies like Ample Communications Inc. to solve Ethernet oversubscription, shows that Daines' view does not always take into account particular bottlenecks between the enterprise network and the edge of the carrier net.

Carriers also have to worry about the "hard guarantees" that are often mandated in service-level agreements. SLAs put strict bounds on latency, jitter, service availability and price-per-packet, and VoIP customers have become adept at specifying voice-service SLAs based on circuitlike performance. If packets are dropped in a congested network, the resultant breakdown of VoIP service could mean heavy charges to a service provider if some sort of TDM infrastructure is not preserved.

The simplest form of QoS in Ethernet and IP networks was developed under the auspices of Integrated Services (IntServ), which provided strict guarantees for each IP flow, and Differentiated Services (DiffServ), which aggregated similar flows for optimized IP delivery. The former could be handled only within the well-defined boundaries of an enterprise, while the latter has been extended to the wide-area network but has not provided the strict latency boundaries for TDM traffic.

The most important extension for MPLS in the realm of circuit emulation has been the Pseudowire Edge-to-Edge Emulation, or PWE3 protocols, developed for the Internet Engineering Task Force by Luca Martini of Level 3 Communications Inc. Pseudowire defined ways for IP flows to emulate T1 and T3 traffic, as well as frame relay or Ethernet, inside an IP/MPLS-routed network. When Ping Pan of Hammerhead Systems extended Martini's efforts to define pseudowire tunnels for new services, the work took on the nickname "Dry Martini."

Where the original Martini pseudowire work was intended primarily for transparent LAN services and virtual LAN services over the WAN, Dry Martini was expected to map Layer 2 services over optical wavelengths, ATM virtual circuits, Sonet cross-connects or other transport options, without requiring MPLS going out to endpoints. Ironically enough, although the work was an extension to MPLS, it eliminated the need to use Layer 3 MPLS to the client. As a result, Dry Martini became known as Transport MPLS.

More ironies were in store. The most effective hardware platforms for implementing Dry Martini usually used multiple switching fabrics. Mangrove Systems Inc. (Wallingford, Conn.) introduced a system at last year's Supercomm show that had separate packet-switched and TDM-switched backplanes to combine Ethernet and Sonet traffic in a single architecture. At the end of 2004, Hammerhead Systems Inc. introduced the HSX6000 system, which melded ATM and IP traffic through dual switching fabrics.

Unless a TDM channel is preserved end to end, however, the issue of circuit resiliency is not addressed in any version of Martini/pseudowire technology. That's why, at last year's Supercomm, Agere Systems Inc. announced work with British Telecom to implement the BT Unbreakable Access algorithm on its Agere Payload Plus processors. The algorithm sets up an alternative path for IP traffic when a primary path is unavailable, offering Sonet-like protection-switching times of 50 milliseconds or less. The protected paths are subtagged IP flows that are tagged at the level of an MPLS path, a field within an IPv4 or IPv6 header, or a tag in a TCP session manager.

While BT already has made the algorithm an element of its "21st Century Network" project, the required backup paths have defined a new problem for carriers: Promoters of IP say the network can be implemented as a replacement for circuits. But if TCP/IP is not implemented as an overlay to a circuit network, hard guarantees of time-sensitive services require some sort of network backup.

The lack of a free lunch is evident at every level. If Ethernet and IP are to emulate TDM service, then a dual-backplane aggregation box may be required. If a Dry Martini system with multiple switching fabrics is not in the cards, then carriers may want to look at ATCA architectures with iTDM buses. If full synchronization and timing information is necessary, a concept like Zarlink's Timing Over Packet may look feasible.

Deepak Kataria, senior engineering manager at Agere, said there are no surprises in such a panoply of choices. Strict timing guarantees in a connectionless network will have to be made in a way that adds physical-layer TDM support or complex Layer 3 support to a simple packet-switched network. At last week's Supercomm, Agere demonstrated packet and circuit-emulated networks that were heavy on pseudowire, but showed a variety of aggregation methods at Layers 2 and 3.

"There will never be one method of T1 bonding, Layer 2 tunneling or TDM buses that completely takes over the network," said Kataria. "That is why programmability in architectures is important-we need to be able to support the services chosen by service providers for different types of networks."