Design Feature: March 3, 1994

How quickly ATM assumes the dominant role its proponents predict is still a matter of opinion, but most companies in the telecommunications industry agree that it will begin its ascent this year. Several key semiconductor components for ATM are already available (Table 1), and many more will make their appearance this year. Most companies in the data-communications and communications-semiconductor industries have ATM design efforts under way, as do the telephone companies. The communications market's needs are forcing them to it (see box, "Why ATM?").

You can best understand ATM's operation by following the path data takes through the network. The broadband-integrated-services-digital-network (BISDN) model (Fig 1) provides a map describing points along the way. Similar to the OSI model used in networking, the BISDN model defines a variety of layers that contain isolated functions sharing data in a fixed format with the layers above and below them.

The BISDN model uses three layers to describe the network-supporting user applications: the physical layer (PHY), the ATM layer, and the ATM-adaptation layer (AAL). PHY handles such details as packaging of cells for transmission, moving data across the transmission medium, recovering data, and determining cell boundaries (framing) within a bit stream. ATM directs the data through the nodes, handles switching, and defines header contents. AAL adapts user data to ATM format and provides data error detection and recovery. A management function operates across all layers, shown in Fig 1 as a plane behind them. The layers include both signaling and control functions as well as data transport and exchange in a fixed format. System decisions at one layer have little impact on the options available at another layer.

Your data's trip through the ATM network begins at the AAL, which comes into play only at the data's source and destination. The AAL is responsible for mapping the user's data into ATM cells. That mapping varies with the quality and type of data-transport service your data needs, adapting the network's apparent behavior to the data's needs rather than adapting the data to fit the network.

AAL has two sublayers: the convergence sublayer (CS) and the segmentation and reassembly (SAR) sublayer. The function of the AAL depends on the class of service you need. That class, in turn, depends on three factors: constant vs variable bit rate, the presence of a timing relationship in the data, and a connection-oriented vs a connectionless relationship between sender and receiver. To date, the ATM specifications identify four classes of service based on these factors (Fig 2). Each class has an associated AAL process.

The first process, AAL1, provides constant-bit-rate services. Such services mimic the behavior of an existing telephone system in that they guarantee a fixed number of ATM cells per unit time to the application. Constant-bit-rate services inherently include a timing relationship. They are also connection-oriented in that they must maintain a connection for relatively long time periods rather than es-tablish a new link each time they have information to send.

The AAL2 process provides a variable-bit-rate service that includes a timing relationship. One application that could use this service is the transmission of compressed video. The video information must arrive at its destination in a timely manner but tends to come in bursts. Users could employ the AAL1 process to transmit this information by using buffering to smooth out the data bursts. But the AAL1 process would cause increased latency, the risk of buffer overflow, and the added cost of buffer memory. The AAL2 process handles compressed video in its raw form.

The AAL3 and AAL4 processes service data-transmission applications that have a variable bit rate with no timing requirement. The AAL3 process works with applications, such as large file transfers and system backups, that need to maintain a link for relatively long time periods. The AAL4 process works in applications that produce short, infrequent bursts of data and don't require the expense of maintaining a virtual channel between bursts. Both processes share a common SAR sublayer that allows for multiplexing of user data streams onto the same virtual channel and provides error detection on individual cells.

A fifth process, AAL5, evolved in response to user concerns that the AAL3 and AAL4 processes had too much overhead. (They use 8% of a cell's data field to provide multiplexing tags and error checking.) The AAL5 process offers no error checking and supports only a single data stream per channel but devotes the entire cell data field to user data.

Regardless of the service supported, the AAL processes have a common function: to break user data into ATM cell payloads and to reassemble them upon receipt. Fig 3 shows a generic AAL operation. The CS breaks user data into intermediate-sized blocks called CS-protocol data units (PDUs). These blocks may include headers or trailers that contain non user information that the receiving AAL process needs for reassembly or error correction. The SAR sublayer breaks each CS-PDU into SAR-PDUs. The SAR-PDU is always 48 octets (ordered sequences of 8 bits) long and is the payload portion of the ATM cell created at the ATM layer. Both processes run in reverse to re-create the user data at the receiving end.

The exact operation of the two sublayers varies with the service provided. In an AAL3 or AAL4 system, for example, the CS-PDU has a 64-kbyte maximum length and contains both a header and trailer in addition to user data. When the CS-PDU segments further at the SAR sublayer, more system information attaches to each cell. By contrast, the AAL5 process puts no limits on the length of its CS-PDU and adds only a few trailer octets to the user data plus padding to make the total a multiple of 48 octets. Its SAR-PDU step adds no overhead at all.

Because it adds no overhead, the AAL5 process makes maximum use of the network bandwidth its ATM cells consume. However, it neither provides error detection nor allows error recovery if a cell gets damaged in transmission. Instead, the application using the AAL5 service must perform any error detection or correction activity without help from the service. The service's inability to determine which, if any, cells are missing further complicates error detection and recovery.

The AAL3 and AAL4 services, in contrast, provide cyclic-redundancy-checking (CRC) error detection and a sequence identifier for each cell. These additions enable a system using these services to detect both 64-kbyte-block and single-cell errors and request retransmission of damaged cells. These features also enable an application to properly reassemble the now-fragmented data stream. The penalty is a 10% bandwidth overhead.

Once the AAL divides user data into 48-octet units, the ATM layer adds the headers to convert the PDUs into ATM cells. The ATM layer then introduces those ATM cells through a switch matrix into the network. Fig 4 shows the content of an ATM cell: 53 octets, of which 48 are data and 5 are a header.

The cell's header has several fields containing information that the network needs to properly handle the cell. The as-yet-undefined generic flow-control (GFC) field will control the flow of traffic from the user into the network at the data's source. At intermediate nodes, additional virtual-path-identifier (VPI) bits replace this field. The VPI and virtual-channel-identifier (VCI) fields tell the network nodes which route the cell will take. The payload-type (PT) indicator shows which type of data—user or control—the data field contains. The cell-loss-priority (CLP) indicator informs the nodes which cells the nodes can discard in the event of congestion. The header error-control bits provide a simple CRC error inspection to allow nodes to discard cells with invalid addressing data.

The ATM layer introduces cells to the network through a switch matrix that allows ATM to efficiently use the network's physical resources. The matrix accepts cells originating in the node's overlaying AAL or from another network node through the underlying PHY layer. Regardless of source, the ATM layer's hardware examines the header of each cell that enters and takes action.

Based on the identifiers, ATM-layer switching hardware typically affixes a routing tag to an entering cell and passes the cell to a self-routing n3n switch matrix, from which the cell automatically emerges, stripped of its tag, at the appropriate output port. The hardware may also rewrite the identifiers so that the next node can use them. All the activity on a given cell occurs before the next cell arrives, so that the switching operation does not cause a backlog in the traffic flow. Some backlog may occur at the switches, however, if an individual switch within the matrix simultaneously receives two cells. To prevent data loss during such a collision, the matrixes typically offer input buffers to hold data temporarily.

Whether the data at a node enters the ATM layer from the AAL or comes through the switch matrix from another node, the ATM layer passes the data to the PHY layer for bundling onto a physical link. The PHY layer has two sublayers: the transmission-convergence (TC) sublayer and the physical-medium-dependent (PMD) sublayer. The selection of a transmission medium determines the operation of both layers.

The ATM Forum, a group of more than 350 companies that are collaborating to set ATM standards, has selected a number of media suitable for carrying ATM traffic, all borrowed from other telecommunications standards. Some of the more important standards include DS3, synchronous optical network (SONET), STS levels 3c and 12c, and 100-Mbps optical fiber using 4B/5B encoding. The ATM Forum is also considering other transmission media, including unshielded-twisted-pair (UTP) wire.

The PMD sublayer for each medium defines such things as wave shapes, bit transmission, framing, and clock recovery. It matches existing standards, reflecting the intent that ATM work within the telecommunications infrastructure. But the TC sublayer is the key to an ATM cell's ability to travel freely on a wide variety of media.

The TC sublayer packages outgoing ATM cells into the medium's framing structure, filling in with null cells as needed. When receiving data, the TC sublayer determines cell boundaries, extracts cells from the bit stream, and checks for header errors. If the cell is error-free, the TC sublayer passes the cell to the ATM layer. The TC sublayer discards cells with errors and null cells (which have a reserved VPI/VCI code).

While a data cell is en route, it passes through the physical and ATM layers at each node. It may change physical media at each node, and the nodes may repeatedly bundle and unbundle the cell. Only at its final destination does the cell pass into the AAL, which reassembles the cell into user information.

The differing AAL services and varied physical layers that ATM offers demonstrate the technology's primary appeal—flexibility. The network's theoretical abilities—to connect any two users over any combination of media, to handle any data type, and to provide only the bandwidth needed—are the features firing the enthusiasm of the telecommunications industry. Visionaries foresee a global ATM network that gives any user access to any other without regard to the link between them. Yet, manufacturers must resolve many issues before they can construct such a network.

Some issues of immediate concern to the industry include traffic management over the network, methods for establishing dynamic virtual connections, and the bridging of ATM to traditional LAN services. The first two issues affect the use of ATM to create a WAN. The third is of more concern to today's potential ATM users.

Traffic management—the handling of congestion at switches, the gathering of traffic statistics, and control of traffic flow—needs definition. In addition, system vendors want to embed such activities in the ATM-layer hardware, which the first-generation ATM ICs fail to do. ATM vendors have targeted the existing hardware at end-user network interfaces more than at intermediate switching nodes.

The methods for efficiently establishing short-term links also need definition. The standards to date address only the creation of permanent virtual circuits, those that users will employ over time periods of hours to months for repeated or massive data transfer. Temporary, short-term transfers need more dynamic, quickly established circuits, so that call origination doesn't dominate the data-transfer time. Short-term transfers must establish efficient circuits, so that the short-term links can be the primary users of the system bandwidth.

Both the traffic-management and short-term-link-creation concerns address needs that arise when ATM forms a WAN, a condition that is still many years away. In anticipation of such a network's creation and to meet their mixed-data needs, early adopters are looking to link their LANs with ATM backbones or to create LANs that are ATM from the ground up. Such systems require that ATM provide "LAN emulation," a function still under definition.

The need for LAN emulation arises because ATM is connection-oriented: It needs to define a link before data flow can begin. Traditional LANs, however, are connectionless. They assume a shared medium and begin data transmission as soon as they have media access. A bridge between the two would have to buffer the LAN message, interpret the address field, and establish the link before beginning retransmission of the data. Because no standards for this action exist, today's ATM systems use proprietary methods. Fortunately, LAN emulation is a high-level software function, and vendors can retrofit systems to comply with standards when they become available.

Despite these concerns and the incomplete state of standards definition, the building blocks for ATM are becoming available for the local level and soon will become available for wider area applications. Once considered a technology of the future, ATM is now taking its first step into today.