DSL modem pooling reconsidered: Modem pooling, or any other means of access concentration, may be inappropriate for DSL

Because it offers a high-speed connection over twisted copper pair, asymmetrical digital subscriber line has gained a great deal of attention as a means of delivering Internet access and multimedia services. As in the case of voiceband modem technology, an ADSL line requires two modems-one on the user end and the other on the network end.

A logical step in ADSL development might appear to be modem pooling-a technique by which one modem on the network end can be shared by many users, with only one user connected at any specific time. This technique is commonly used for voiceband modems in telephone networks with a line concentration of 6-to-1, 8-to-1 or, in some cases, 10-to-1, resulting in low-cost service deployment. In ADSL vernacular, such proposed systems are also referred to as digital access multiplexers, line concentrators or dynamic allocation matrices.

Modem pooling may be inappropriate for ADSL systems, however. It may cost more to implement than it saves in modem costs. And unlike conventional analog modems, which customers can use to connect to multiple modem pools, an ADSL modem pool resides in the central office, through which all calls must pass. This creates a potential bottleneck and reduces flexibility.

Modems and line concentration ADSL offers speeds up to 8 Mb/s downstream and up to 2 Mb/s upstream. Actual bandwidth depends on line condition, distance and service configuration. A low-pass filter separates frequencies used for POTS, thus enabling a copper pair to be used for both telephone service and ADSL-based high-speed data service at the same time.

The reference model developed by the ADSL Forum and the ATM Forum uses the term ADSL transmission unit (ATU) for the ADSL modem (Figure 1). The ATU at the CO or network end is referred to as the ATU-C, and the modem at the remote or customer end is referred to as the ATU-R.

When sharing telephone service on the same copper pair as ADSL service, the POTS-C and POTS-R interfaces refer to the connections to the public telephone network and home wiring, respectively. An access node is a network element that aggregates high-speed data from many ADSL lines-and possibly other access nodes-onto a single network interface to the backbone network.

A telephone company's access network connects customer lines to a CO switching system, remote switching system, universal digital loop carrier (UDLC) system or integrated digital loop carrier (IDLC) system.

In the case of a CO switching system or a remote switching system, a customer line terminates on the main distribution frame (MDF) and then on a line unit that is part of the switching system. In a UDLC system, customer lines terminate on a remote terminal that is connected to a CO terminal via a T-1 interface, and the CO terminal connects to a line unit on an individual customer basis. An IDLC system terminates customer lines on a remote terminal and offers a multiplexed T-1 or Sonet interface to an integrated digital terminal that is part of a switching system.

In an ADSL system, the ATU-C would be collocated with the MDF of a CO or remote switching system or with a channel unit of a remote terminal in a UDLC or IDLC system.

Line concentration generally refers to the ratio of subscribers to resources provided by the network. A distinction should be made between concentration of network resources and concentration of access equipment.

For direct termination on a line unit, the telephone switching network is engineered-based on industry guidelines-to achieve a 6-to-1, 8-to-1 or 10-to-1 line concentration of access equipment. In DLC systems, line concentration is achieved by provisioning bandwidth on the remote terminal connection to the CO. No concentration is built into the access equipment. The critical network resource-the switching system time-slot interchange-is efficiently managed through bandwidth management on the distribution network.

A typical Internet service provider connects to its local telephone company CO via a T-1 or primary rate ISDN line (Figure 2). A modem hub is located at the ISP site, and customer calls are routed to it based on the dialed digits-similar to the routing of voice calls in a PBX environment.

Customers can connect using analog modems between 14.4 kb/s and 56 kb/s and are limited by the 64 kb/s throughput of the core switching network. In the case of basic rate ISDN service, two voice channels are banded together to provide a 128 kb/s connection.

An ISP usually engineers its network with one modem on the network side for every 10 subscribers in the serving area. The core switching network provides access to the customer, whereas the ISP network provides the additional service of modem support.

Regardless of the availability of a modem at an ISP location, a customer can dial in to the core network-which recognizes a modem's busy tone and allows customers to dial into a different modem pool or use the line for other purposes.

In terms of ADSL systems, a distinction is made between a network access provider (NAP) and a network service provider (NSP). The local telco is the NAP, whereas the ISP would be the NSP.

A typical analog modem deployment architecture has many NSP locations within reach of one NAP. This allows a customer to dial into alternative modem hubs if the primary location is busy. Although an ISP location may be engineered for a 10-to-1 concentration of customers to available modem resources,

customers, in effect, can achieve a 1-to-1 concentration ratio because they have the flexibility of connecting to modem hubs in multiple locations.

Line concentration in telephone networks and modem concentration in ISP networks primarily are based on four traffic characteristics of the system-inter-arrival time, call holding time, serving area density and desired quality of service.

In voice applications, inter-arrival time-the time between calls-is large. The typical customer places or receives three to seven calls per day. Call holding time is short, with the average conversation lasting just three to five minutes. And the density of the serving area typically allows about an 8-to-1 concentration. Using these parameters, quality of service is unaffected by network engineering. Once a connection is established between the calling and called parties, network bandwidth is dedicated for the call's duration.

In data applications inter-arrival time is small. The average user connects 15 to 20 times per day, counting separate instances of file server access, printers and e-mail. Call holding time is long-often measured in increments of 10 minutes to hours. Because of these different traffic characteristics, telcos often have to re-engineer COs serving ISPs to maintain quality of service.

Traditional voice networks are engineered with concentration in access equipment and no further efficiencies are derived in the backbone network. Data networks are engineered in just the opposite manner. Although access equipment has no concentration, data networks derive significant benefits in bandwidth use. A business complex of 1000 users with three to five 100 Mb/s local area networks, for example, may derive adequate performance from a single T-1 connection to the backbone network.

However, line concentration, a cost-effective design for telephone networks, may be inappropriate for data networks.

ADSL peculiarities When ADSL modems are used to provide Internet access, the modem typically terminates at the NAP, and data from many ADSL lines is multiplexed onto a single network interface for further connections to multiple NSPs (Figure 3). Internet access would be one of the services offered by an ADSL access network, and an ISP would be one of the NSPs connecting to the access network.

An ADSL access network can connect to multiple ISP networks, thus allowing customers a choice of service providers, or to corporate networks for telecommuting and at-home workers.

Recent discussions of ADSL deployment have focused on pooling ADSL modems-which is the equivalent of line concentration of ADSL lines-under the concept of a "dial-up ADSL network."

Dial-up ADSL is intended to enable access providers to concentrate multiple customer lines onto a pool of limited ATU-Cs, where the number of ATU-Cs is less than the number of subscriber lines, or ATU-Rs. It relies on establishing of a connection on demand between the ATU-R and ATU-C.

A dial-up ADSL architecture requires new equipment (Figure 4).

The ATU-R should be able to generate a "request for service" tone. A tone detector is necessary for each customer line to receive connection request and line code information from the corresponding ATU-R and to convey the request to the switch controller. A switch controller controls the concentrating switch matrix, and the switch matrix connects active ATU-Rs with ATU-Cs.

The connection request is initiated by generating "request for service" tones. When a 1-to-1 ratio of ATU-Cs and ATU-Rs is used, either end may initiate this request. When modem pooling is used, however, only the ATU-R can initiate a connection because each ATU-C must serve multiple ATU-Rs.

Customers can access a network-based application in three ways-through a user-initiated connection request, an application-initiated connection request or an ATU-R-initiated connection request.

The first method is the most intuitive. Customers would use their computer to log onto a remote network just as they use analog modem pools today. The other two methods are transparent to the user but are not uncommon.

In the application-initiated connection request, the application that uses or intends to use a connection can request the set-up and tear- down of the connection. For example, a customer might send an e-mail message through an e-mail interface, which would automatically initiate the required modem connection. This scheme, however, violates prevailing precedence rules: In networking systems, the link connection and the end-to-end connection should be established before an application starts to use the connection.

The ATU-R-initiated connection request is used in modem pooling applications to compensate for the fact that the ATU-C cannot initiate such a request.

In this scenario, the ATU-R requests connection set-up when it detects the presence of upstream data-such as an e-mail server advising the end user that a message has arrived. When the ATU-R detects absence of traffic for a given amount of time, it tears down the connection.

This, too, violates prevailing precedence rules. Once the ATU-R tears down a link connection, and after it has sensed silence for a period of time, end-to-end connections that use the link connection will be lost and applications eventually will abort.

Many studies focusing on data communication and Internet navigation have revealed that users often experience long holding times while reading certain information or just staring at the screen. Such users connecting through an ADSL modem pool would be automatically disconnected so that the ATU-C can be made available for other users.

The user may or may not detect the application status until the next time he requests information or invokes an action. At that point, he has no choice but to reboot the application and reinitiate the connection. These events are due not only to the way modem pooling works, but also to the way software applications work. Unless all parties involved make a concerted effort to correct this, the problem will remain.

Another disadvantage of ADSL modem pooling relates to its dependence on the physical layer.

Did you ever wonder why 28.8 or 33.6 kb/s modems can interoperate, but the two competing standards for 56 kb/s modems-X2 and Kflex-cannot connect seamlessly to each other? The answer lies in the physical layer.

Modems achieve a certain "speed" through algorithm design, encoding and decoding of data, error detection and correction techniques, loop length, and loop conditioning. The modem on the user end has to match the configuration of a modem on the network end for ADSL systems to work together. In a modem pooling design, the ATU-R and ATU-C must be trained and synchronized each time the ATU-R requests a connection.

Currently, ADSL modems using various line coding methods-including carrierless amplitude/phase (CAP) modulation, quadrature amplitude phase modulation (QAM) and discrete multitone (DMT)-cannot talk to each other. And modems from different vendors using the same line coding method may even need some tweaking before they can interoperate properly-an effort the industry is currently undertaking.

The maximum data that can be offered to any given ADSL modem user depends on loop length, loop condition and other services present in the network, such as high bit-rate DSL and private-line service.

According to its definition, a rate-adaptive DSL modem is designed to automatically determine the maximum permissible throughput for any given user line. The service offered to a given user defines the modem configuration and, subsequently, the ATU-R and ATU-C are set to that configuration. Any changes to this configuration under current product conditions require a "modem retraining" in which the user and network modems are adjusted to a new set of parameters. The process takes several seconds and may require an initialization of the physical layer.

In order for "modem pooling" to work, all the modems must have a homogeneous configuration so they do not have to go through "retraining." Given different loop lengths, loop conditioning and user preference to select different grades of service, imposing a uniform configuration on all modems is unrealistic. Managing different classes of service via operational procedures could outweigh any savings that may be realized due to modem pool design.

In current ADSL deployment architecture, the cost of network routing-as incurred in backbone network and gateway routers-far outweighs the cost of access equipment. With current trends in modem design and silicon integration, it is likely that the same trend will continue in the future.

Unlike voiceband modems, a busy tone on an ADSL modem prevents a customer from connecting to any network service.

New traffic and plant measurements will be required in order to facilitate proper engineering and administration of an ADSL access network. And because subscriber quality of service offering is closely tied to ATU-R and ATU-C modem configurations-both being set to the same data rates-additional cost is incurred in engineering CAP, DMT or QAM modem types for each service type.

In the long run, the additional operations costs may exceed the capital gains because of modem pooling architecture.