Here Comes UBR +: The ATM Forum's guaranteed frame rate initiative paves the way for potential new revenue for service providers and affordable guarantees for users

The ability to elegantly support quality-of-service guarantees remains a major differentiation point for asynchronous transfer mode among high-speed networking protocols.

Fundamentally, ATM strives to integrate traffic from disparate sources over a common backbone infrastructure. The nature of the offered traffic can range from standard voice to video and Internet protocol and various types of local area network data. Whereas optimized networking protocols already have been developed to efficiently handle many of these traffic types individually, ATM looks at ways to bring all of this traffic together.

By leveraging ATM, service providers can offer customers one-stop delivery for a variety of services by providing frame relay, Internet access, native ATM, asymmetrical digital subscriber line, and backhaul of aggregated voice over a single infrastructure with unified billing and maintenance. This network consolidation results in significant operational cost savings that can be passed on to the user.

Stratification of offered traffic based on cell delay and cell loss requirements-which vary depending on the application-can offer flexibility to service providers. By leveraging ATM quality of service (QOS), service providers can offer differentiated data services.

Service providers could offer "gold" service to premium users to support telecommuting and home office applications, as well as "silver" or "bronze" service for casual Internet users. Service could be valued anywhere from less than $5 a month to more than $50 a month depending on the customer's performance requirements.

Additionally, service providers can use network resources more efficiently by better "packing" backbone links with statistically multiplexed lower priority traffic, leading to maximized service revenues.

The appeal of ATM QOS is driving a new class of service-guaranteed frame rate (GFR)-which is currently being defined by the ATM Forum to address the skyrocketing demand for cost-

effective Internet access and the significant volume of TCP/IP traffic that is offered to the service provider network.

It will join several existing ATM classes of service, including:

* Constant bit rate (CBR), which handles real-time, constant bandwidth traffic such as voice, with strict guarantees for cell loss, delay and jitter.

* Real-time variable bit rate (VBR-rt), which targets video and delay-sensitive data and also provides loss and delay guarantees.

* Non-real time variable bit rate (VBR-nrt), which handles bursty data traffic that can tolerate some delay but offers dedicated virtual bandwidth.

* Available bit rate (ABR), which is optimized to handle typical IP and legacy LAN traffic and provides a service guarantee based on a minimum cell rate. It depends on ATM layer flow control to throttle sources during congestion and avoid overloading the core network.

* Unspecified bit rate (UBR), also optimized to handle typical IP and legacy LAN traffic but without service guarantees.

From the service provider's perspective, a more interesting treatment of the different classes of service lies in their relative "cost."

Although service differentiation comprises many components-such as bundling of applications, service, reliability and billing/call detail record options-"bandwidth" is the major commodity being offered by the service provider to its customers. The goal is to maximize revenues by effectively provisioning available bandwidth to serve as many customers as possible while maintaining the expected QOS.

In general, as the desired level of guaranteed service increases, the bandwidth allocation flexibility of the service provider decreases (Figure 1).

Controlling QOS The algorithms that determine the network's capacity to accept an incoming call are referred to as connection admission control. The connection admission control algorithm continually monitors available network resources and determines, based on an incoming call's traffic descriptors, whether the call can be handled without sacrificing the QOS of calls already admitted to the network. On this basis, calls are accepted or rejected.

In this light, the real-time classes of service-CBR and VBR-rt-are clearly the most taxing on the available network resources. Performance is guaranteed within tight parameters for both cell loss and cell delay.

Engineering guidelines for CBR traffic are the most straightforward because no statistical multiplexing for CBR connections exists. The peak cell rate and the sustainable cell rate are identical. The user pays for the allotted bandwidth for the duration of the call regardless of whether or not traffic is injected into the network. This model is similar to capacity provisioning in a time division multiplexed environment.

For traffic that is not critically delay sensitive, ATM draws upon statistical multiplexing to allow many users to dynamically share pools of bandwidth. This traffic is usually highly bursty in nature. The level of burstiness in a connection is often quantified using a factor-called beta-that represents the ratio of the peak cell rate to sustained cell rate. Typically, the peak cell rate is substantially higher.

For non-real time sources, the connection admission control algorithm must calculate an accurate equivalent bandwidth estimation, which quantifies the bandwidth demands of the incoming call. Based on this estimate, bandwidth is allocated to ensure that the QOS metrics are met for the call.

The service cost is proportional to the level of guarantee that is provided to the customer. Hypothetically, a service provider might tariff a 5 Mb/s CBR circuit at $3000 a month. Alternatively, the provider could allocate the same 5 Mb/s of network resources among 10

VBR-nrt subscribers with a peak cell rate of 5 Mb/s and a sustainable cell rate of 0.5 Mb/s at a cost of $500 each per month. And perhaps as many as 200 UBR subscribers could be accommodated in the same amount of bandwidth at a cost of $30 a month each (Table 1).

Herein lies the potential for leveraging ATM statistical multiplexing to maximize service revenues while leveraging ATM QOS to segregate traffic with varying requirements and guarantees. UBR services offer the lowest-cost services available in ATM solely because they extend no performance guarantee. The network accepts all UBR traffic, and serves it as bandwidth becomes available. Because no QOS guarantee is provided, the connection admission control algorithm is not performed on UBR connections.

The UBR class serves to "pack" the transmission links to their engineering capacity and allows service providers to effectively maximize revenue by adding low-priority customers without jeopardizing the QOS of higher-paying customers with service guarantees. The UBR class of service is potentially the best alternative for offering low-cost data services such as Internet/intranet access and non-critical data transfer.

However, customers tend to be wary of a public service offering that extends no throughput guarantee whatsoever. This is especially true as frame relay increasingly becomes a commodity service. With the rapid expansion of frame relay services and ATM-based backbones for backhauling frame relay traffic, some level of guaranteed QOS is highly desirable. For non-casual subscribers, the "send-and-pray" nature of UBR service reduces its appeal regardless of cost.

Bridging the gap The new GFR initiative attempts to strike a balance by offering customers a low-cost ATM service class with a performance guarantee.

The class of service also is referred to as UBR with a minimum cell rate, or UBR+. The notion of a minimum cell rate is borrowed from ABR, which defines a minimum throughput guarantee for the connection. The intent is to keep the minimum cell rate relatively low, thereby reducing the cost of the offered service.

ABR specifies the behavior for end systems-virtual sources and virtual destinations-which provide the intelligence to pace their injection of traffic to the network based on signaled feedback. ABR pushes congestion to the edge of the network.

The intelligence of ABR is quite powerful but carries with it the complexity and cost of intelligent adapters at the edge, which do not cost-effectively exist today. Interoperability is another major issue in offering end-to-end ABR flow control.

UBR+, however, does not include the sophisticated closed-loop flow control specified by ABR. In this manner, UBR+ bridges the gap between ABR and UBR service categories by offering a guaranteed data service within a reasonable cost structure.

The statistical nature of data traffic typically causes the actual throughput of a UBR+ connection at any instant to be much higher than the minimum cell rate, but the minimum cell rate serves to provide the QOS guarantee. Adequate buffering is required within the ATM nodes to accommodate the highly bursty UBR+/UBR traffic as it awaits servicing behind the higher-priority classes of service (Figure 2).

Algorithms such as round robin and exhaustive round robin are commonly used to service queues of prioritized traffic within a switch. The round robin algorithm alternately serves queues without regard for traffic priority. The exhaustive round robin algorithm serves the highest-priority queue until empty before proceeding to the next highest-priority queue. It further pre-empts serving lower-priority queues upon arrival of a cell into any higher-priority queue.

ATM developers have devised sophisticated buffering and queue servicing methods to address priority and fairness issues within ATM. Without such methods, lower-priority traffic potentially could wait in the queue indefinitely-a phenomenon known as "starvation."

Developers have modified the round robin and exhaustive round robin algorithms to accommodate differentiated data services and prevent starvation of the UBR/UBR+ queues behind CBR and VBR traffic. One efficient mechanism is to serve the real-time queues using an exhaustive round robin algorithm and to serve the non-real time queues using a queue-length weighted round robin. The final arbitration between the two queues is also via an exhaustive round robin. This scheme is very effective for carriers offering services that span CBR, VBR, ABR, UBR and UBR+.

Within the proposed ATM Forum specification for UBR+, both the minimum cell rate and peak cell rate may be specified on a per-connection basis. Ingress policing is performed on the peak cell rate, with discarding of non-conforming cells. More important, connection admission control is performed on the minimum cell rate of each connection (Figure 3). Thus, the minimum cell rate is guaranteed on the aggregate of UBR+ connections, so a theoretical upper bound exists on transit delay of UBR+ packets.

In order for a UBR+ service to prove viable, the traffic management scheme needs to include intelligent packet discard mechanisms that increase the effective throughput of the transmission links within the network. In dealing with TCP/IP traffic, or traffic from any protocol with inherent higher-level flow control mechanisms, it is critical to realize that upon reassembly at the destination, a retransmission of the entire packet occurs if even a single cell of the data packet has been lost.

Mechanisms such as early packet discard serve to drop entire packets as congestion nears, in an attempt to avoid randomly dropping a few cells from several different packets at the onset of congestion. Intelligent packet discard methods such as these increase the effective throughput of the links by greatly reducing the number of "dead" cells that are carried in the network. Dead cells are those that ultimately will have to be re-sent because of a lost cell elsewhere in the packet. The measure of effective throughput is commonly referred to as "goodput," and this notion lies at the heart of ensuring good performance for IP traffic riding in the UBR+/UBR/ABR classes of service.

Using UBR+, the end user can trade off delay vs. service cost. For applications that can tolerate delay-such as TCP/IP applications, non-real time file transfers, and Internet access-the user can select a very low minimum cell rate.

By offering customers a UBR+ cell relay alternative that allows modest guaranteed throughput at a low cost, service providers can maximize use of network resources and realize increased revenues while attracting new customers to a viable and affordable data service offering.