Fly my way: ATM's efficiency challenge: traffic management as a key differentiation for QOS and SLA guarantees

Asynchronous transfer mode in the wide area is taking off. By integrating diverse traffic types, a multiservice ATM network offers compelling economic advantages over application-specific networks.

Yet most wide area ATM deployments have focused on data-only applications in support of frame relay traffic. In some cases, limited voice and video traffic has been introduced.

But data, voice and video have very different characteristics. Voice and video traffic have strict network delay and loss thresholds, collectively described as quality of service (QOS). To meet these thresholds, some network managers have had to overprovision bandwidth, thereby reducing efficiency.

Throwing bandwidth at the QOS problem prevents ATM from delivering the efficiency that motivates carriers to deploy ATM in the first place. ATM was designed for multiservice applications, and this is clearly the direction for the intelligent ATM shopper. But the real trick is to provide the QOS required by disparate traffic types such as voice, video and data while sustaining high usage.

This may sound easy but not for the ATM switches now deployed in the market. These early switches lack advanced traffic management techniques required to meet the challenge. These new techniques, which provide higher link efficiency and QOS guarantees simultaneously, are the key that separates the emerging ATM switching platforms from existing ones.

Transporting traffic over an ATM wide area network has many parallels to operating an airline. Everyone is happy if prices are reasonable, the planes are full, they take off and land on time, and passengers are delivered safely. Airlines need effective reservation systems and air traffic control to operate efficiently and deliver the service that customers expect.

Similarly, ATM WANs need effective advanced traffic management to operate efficiently and deliver QOS requirements that customers expect. Congestion delays and cell loss can lead to retransmission, causing additional network congestion. Improperly managed, a retransmission "storm" escalates into a hurricane, potentially crippling the network.

Efficient takeoffs Switch architecture has a profound effect on the network's ability to sustain high usage and simultaneously deliver QOS requirements. Contention points in a switching fabric impede high network usage and degrade QOS by causing delay and cell loss.

A contentionless switching fabric lets traffic enter and pass through the switch unimpeded-or without multiple intermediate buffering stages, head-of-line blocking and the associated delay and cell loss.

Did you ever sit in a plane at the gate, waiting to take off because of air traffic congestion? The plane is being held up at the takeoff point just like network traffic gets held up at contention points. In much the same way, traditional switching fabrics degrade QOS and efficiency in an ATM switch.

Today's ATM WAN switches cannot pass cells from input ports to a single output port without multiple contention points. This is a basic architectural issue, just as airports with one runway cannot handle as much traffic as those with multiple runways.

It is contention-at the input ports, within the switch fabric and at the output ports-that creates problems. Contention points introduce delay, cell loss and potential escalation into congestion and performance degradation. Managing contention requires that advanced buffer management mechanisms be employed for more complexity, but the switches deployed today are not equipped for this. The result is poor network usage and degraded QOS-not exactly an incentive for deploying ATM technology (Figure 1).

A contentionless architecture provides dedicated parallel data paths through the switch, avoiding delay and loss. Contentionless switch fabrics avoid head-of-line blocking, back pressure overhead and other problems associated with multiple contention points.

These problems can result in annoying gaps or skips in video and voice connections, and impaired QOS and customer satisfaction. Employing an ATM switch with a contentionless switch fabric is like having a private runway-regardless of what happens in the rest of the airport, your plane can take off and land without delay.

Effective air traffic control When cells traveling to an output port exceed the port's transmission capacity, output port buffers provide temporary storage. To maintain the appropriate QOS for each connection, the switch must efficiently manage output port buffers by prioritizing and allocating buffer and bandwidth resources, or QOS and network efficiency will be compromised. This is similar to airport traffic control's job managing the runways.

For effective QOS levels, the output port buffer must make intelligent allocation decisions based on service classes, such as guaranteed QOS classes like constant bit-rate, near real-time variable bit-rate and real-time variable bit-rate, and best-effort service classes such as autobaud rate detect and unspecified bit-rate.

For guaranteed service classes such as voice and video, bandwidth and buffer space must always be available when required. Like first-class passengers with reserved seats, first-class network traffic gets priority. Best-effort classes contend for a common pool of buffer and bandwidth left by connections in the guaranteed service classes, much like standby passengers.

Intelligent buffer allocation decisions must deliver three things: negotiated QOS, fairness among virtual circuits and efficient network usage.

Today, most ATM switches implement a static discard algorithm-based on set discard thresholds-either for each service class or for each connection. Both schemes allow differentiated QOS on a per-class basis. But the first scheme fails to achieve fairness among connections in the same service class, while the other fails to provide efficient use of the available buffer and bandwidth. So another approach is necessary-per virtual circuit accounting capability.

Reservations, service and arrival time Per virtual circuit accounting capability allows an ATM switch to meet the goals of fairness, QOS guarantees and efficiency simultaneously. It does so by allocating per virtual circuit buffer and bandwidth space dynamically in response to real-time conditions by considering two factors: the number of cells buffered for each virtual circuit and the instantaneous number of cells buffered for the corresponding class of service (Figure 2).

The relationship between these two factors can be illustrated by a curve. Based on instantaneous demands on the switch, each virtual circuit is independently determined to be above or below the curve, thus providing or denying additional buffer space.

Performing the test on a per virtual circuit level provides fairness. The end result-bandwidth hungry, non-guaranteed services get bandwidth when it's available but are throttled back fairly when bandwidth is scarce.

Key passenger for ATM transport In an ATM WAN, fairness can mean the difference between an application working well or not working at all. Consider new Internet protocol applications such as streaming voice and video. They employ user datagram protocol ( UDP) rather than TCP.

Unlike TCP, UDP does not behave fairly and flow control the source, so even during network overload, UDP traffic keeps flowing, exacerbating the problem, while TCP traffic backs off. During severe congestion, TCP traffic comes to a halt, but UDP packets continue to burst. In short, unfair sharing of resources can cripple the performance of an ATM network if effective traffic management is not employed. To a data service provider or Internet service provider, this could mean that customers experience poor performance.

Fairness is critical, not only among applications but also among individual virtual circuits. Fair access to buffer resources is vital to QOS delivery.

Per virtual circuit accounting in ATM can be compared to an airline enforcing carry-on baggage limits when a flight is full and relaxing the limits when flights are lightly loaded.

Multiple, distinct service classes allow an ATM network to handle a diverse mix of traffic. By supporting a large number of distinct service classes, a service provider can match application requirements to service capabilities closely and efficiently.

A close match between QOS required and QOS delivered reduces the need to overcommit resources, just as some travelers find that business class suits their needs better than first class or coach. And multiple service classes enable differentiated service offerings with various pricing options. Advanced traffic management provides at least 16 service classes.

Switches deployed today typically adopt a strict priority queuing discipline. Cells of a higher traffic class get non pre-emptive priority over cells of a lower one.

The problem with a strict priority discipline is bandwidth starvation. The lower-class traffic suffers from bandwidth starvation when higher-class traffic exceeds the available bandwidth. This problem has the potential to escalate into serious congestion as the source retransmits delayed IP packets.

Advanced traffic management avoids bandwidth starvation by allowing a service provider to assign each service class a minimum bandwidth guarantee. This reduces the QOS effect on lower-priority traffic of the temporary presence of excess higher-priority traffic, just as some discount fare seats are available on all flights.

A common pool of bandwidth can be set aside and shared on a priority basis. Any service class temporarily receiving more traffic than its reserved bandwidth can resort to the common pool of bandwidth for relief.

Traffic shaping Traffic shaping represents yet another opportunity to improve end-to-end network performance. It smooths the bursty cell streams to create a more predictable traffic profile. This results in better fairness, lower cell loss and less stress on network resources.

Shaping can bring non-conforming traffic into conformance with the contracted service level agreements, thereby eliminating the prospect of cell loss caused by policing at other nodes in the network. But to be effective, shaping must be performed on both per virtual circuit and per virtual path bases. ATM switches deployed today shape links only at the virtual path level and can't create conformance to contracts on individual virtual circuits.

Advanced traffic management now emerging provides both.

ATM advanced traffic management delivers efficiency, fairness and differentiated services, all while maintaining QOS requirements. The key factors affecting advanced traffic management are contentionless switch fabric, adaptive buffer allocation, many service classes and per virtual circuit shaping.

With advanced traffic management, ATM WANs can carry diverse traffic types efficiently, delivering the economic advantage that was the reason for deploying ATM in the first place. Whether you are optimizing use of a network or an airline reservation or scheduling system, it helps to operate near capacity, on schedule and with quality service that creates return business.