June 1999

COACH, BUSINESS, OR FIRST CLASS?
Internet QoS

BY SURYA PANDITI

Like the airline industry, with it varied service options, in order for the Internet to become truly business friendly it should provide quality of service (QoS) levels corresponding to different classes of traffic. These classes would allow a range of services customized to suit everything from mission-critical transmissions to Web surfing. Many service providers today are scrambling to improve their overall network performance in order to offer businesses a variety of service types, but this is an ineffective use of resources which can directly impact the cost of providing service and the provider's ability to operate profitably in different market segments. And while some service providers say they provide QoS, they are actually limited to massaging information carried solely on their networks. Currently, there is plenty going on to change this.

THE QoS VALUE
Internet QoS is essential before effective service-level-agreement (SLA) revenue models are possible. Enterprises should be able to buy only the type of service they need to suit their particular business models. Carriers selling the service could audit for proper billing based on traffic, transmission time, or QoS level. Service differentiation would support an array of uses, including VPNs, e-commerce, Web surfing, high-bandwidth allocation, and high-speed transmission for those with business-critical applications or managing LAN and legacy services. Billing, of course, would be appropriate to the service class.

SETTING STANDARDS
Standards committees and communications equipment vendors are crafting the future of QoS over IP. The Internet Engineering Task Force (IETF), a leading industry group, heads the effort to formalize differentiated services (DiffServ) and develop the MultiProtocol Label Switching (MPLS) standard. Both are key to providing QoS throughout the Internet.

While early IETF efforts designed protocols (e.g., Streams-II, RSVP) providing characteristics to discrete end-to-end flows, they didn't provide quality to the overall Internet, which supports millions of concurrent flows, because it's impossible to provide quality on a per-flow basis. For flow-based protocols, this becomes a hydra-like scaling problem. But even while many designers were busy implementing proprietary mechanisms to let customers group or class traffic throughout their own networks, the IETF was focused on providing QoS for all users in the Internet.

Traffic flows need to be consolidated in order to provide QoS for the entire Internet. Traffic must be summarized into aggregate groups or classes throughout the network, in much the same way that routing protocols provide summaries for networks to the Internet's routing table. Recognizing this, last year the IETF formed a working group to formalize DiffServ as an IP QoS mechanism that lets network traffic flow with enough quality to ensure successful application deployment.

The primary component allowing QoS to be deployed with DiffServ is the identification of the traffic class of each packet. Traffic is directed without the overhead of managing each end-to-end flow. The IETF working group provided tagging by re-defining the contents of the "type-of-service" field in both the IPv4 and IPv6 headers. Previously unused, this field has been in IP since its inception.

A differentiated services network can accommodate easily both a new-found "Internet company" that transacts business entirely over the Net or a traditional business migrating from a private voice/data network to an IP-based service provider network. Both can be provided the service levels they demand, as DiffServ affords a nearly unlimited suite of service combinations to be carried over the Internet and the ISP's network.

The IETF also has devised an MPLS standard based on the merging of connectionless protocols like IP with virtual-circuit networking concepts. This is a more efficient way to engineer traffic than using ATM. Without MPLS, there is no way to ensure high-priority traffic won't exceed the physical capacity of any link in the network. MPLS lets a service provider define the physical links that specific traffic will use, making it possible to manage traffic load smaller than the links ability to handle the traffic with acceptable quality.

ADAPTING FOR THE FUTURE
With nothing inherent in TCP/IP protocols to preclude it, QoS over IP will be as easy to implement in TCP/IP as in ATM. In fact, many of the same QoS algorithms that work on 53-byte ATM cells can be leveraged by IP switch/router vendors to work with variable-length IP packets while still able to direct mapping of ATM traffic. Adapting ATM algorithms to TCP/IP means that the QoS end users demand will be deliverable through a standard routed infrastructure. Good examples of such adaptation include Random Early Discard (RED) for congestion management and Weighted Fair Queuing (WFQ) for better traffic prioritization.

Random Early Discard (RED) works with TCP to detect and avoid network congestion by identifying temporary "bursty" traffic and discarding traffic - all of it, if necessary - as the transmit queue becomes larger, rather than waiting for the queues to fill before deciding to "tail drop" packets. This practice allows end-to-end flow control functions like TCP to adapt to constant forwarding rates instead of trying to match the start/stop nature of tail drops.

Different discard profiles are allowed for each class because RED can be independently applied to any traffic class. This means fair bandwidth reduction can be applied across the network in proportion to the network's utilization, with QoS levels maintained for high-priority traffic like SNA, voice, and video.

Weighted Fair Queuing (WFQ) can be employed by carriers in the core network to provide high-priority traffic (i.e., voice or video) with a higher level of service than simple data traffic, with the left-over bandwidth then split among the low-priority requests. To take advantage of this, edge routers and frame relay access devices (FRADs) must present traffic to core switch/routers using IP-precedence bits that request one of six possible classes of service from the network (as well as access control lists to define policies required for transmission through the network). These core network routers can deliver to end users a CIR (committed information rate) across the network by classifying, filtering, and shaping traffic flow using techniques such as CAR (committed access rate) at the network edges.

SUMMARY
QoS is used sparingly today in networks. Just around the corner, however, expect service providers to widely adopt QoS and DiffServ to slice management costs and add service offerings to the revenue stream. Carriers will feel the pressure to adapt backbones to Internet protocol as IP increasingly becomes the foundation for wide-area corporate communications. Visionary enterprises looking to the Internet for success should be pushing their ISPs (and their carriers) to think beyond just maintaining an Internet backbone. Businesses deserve - and will be demanding - a business-class, multi-service IP network optimized for quality voice, video, and data.

Surya Panditi is president and CEO Avici Systems. Based in North Billerica, MA, Avici brings an innovative architecture in support of the IP switching paradigm, delivering up to 40 gigabits/port (OC-768) forwarding rates and multi-terabits of switching capacity, while simultaneously supporting the QoS required by emerging applications. For more information, please visit Avici's Web site, at www.avici.com.