IP telephony, toll quality and the WAN, together for the first time

Over enterprise campus networks, IP telephony is delivering impressive gains in telecommunication management flexibility and significant cost reductions. To extend its benefits over WANs, however, a new approach is needed. Adaptive networking offers considerable promise.

As long as IP telephony applications run on a private campus network under the control of a knowledgeable IT staff, users can expect application availability and voice quality comparable to the POTS network. Management can expect to slash hundreds of thousands of dollars in annual telecom costs and make adds, moves, and changes as easy as unplugging an IP phone, walking it to a new location and plugging it in. Advanced features such as Interactive Voice Response (IVR), unified messaging and multimedia conferencing can be delivered cost-effectively to multiple enterprise locations.

WANs, with their global reach over a shared infrastructure, offer the potential to economically link widely dispersed enterprise locations and effectively deliver data and voice communications. However, as latency-sensitive voice and data applications begin to converge over the WAN, reality hits.

Deploying IP telephony over a WAN yields the best results if private networks are used. Compared to the public IP network, private networks generally offer better application availability and call quality. The main drawback of installing a private network for voice applications is cost. For organizations that operate around the world, intercontinental private links would be cost-prohibitive. And even private networks rarely deliver more than 99.9% or "three nines" availability--not good enough for toll-quality voice which requires an availability of 99.999% or "five nines." With the public IP network, the situation is even worse. While the many networks comprising the Internet may be up and running, in spite of recent progress, the public IP infrastructure rarely exceeds application availability measurements of 99% availability.

Why? Even at its best, the public IP network is fundamentally a "best effort" medium with limited quality of service (QoS) guarantees. Many competing applications with varying transport requirements share the common bandwidth. They are transported over a wide variety of switches, routers, firewalls, load balancers, packet shapers, virtual private network gateways and other devices, all of which have varying software systems, feature sets and maintenance windows. Software compatibility issues, congestion, and other performance-affecting events occur constantly, creating outages and other quality problems. As soon as a voice transmission leaves the enterprise network, so does all control over how it is handled.

Having enough bandwidth is not the problem. All networks suffer quality failures, such as delay spikes, packet loss, link failures, delay fluctuations and jitter. To complicate matters, these problems rarely occur in all networks at same time, making it impossible to predict and compensate for failures. As a result, network performance problems result in application outages that can affect calls for a few seconds or multiple minutes. No network--carrier or enterprise--is immune.

A joint study conducted by Stanford University and RouteScience in June 2001 looked at voice application performance over service provider networks and found the following:

• Delay and delay variability occurred on all network links studied
• Sixty-three percent of the calls experienced unacceptable quality during key periods
• Three percent of the calls never had acceptable quality
• There was propagation delay for calls along the east coast of 3.25-11.8 milliseconds
• Propagation delay for calls from the east coast to Colorado was 28.3-77.8 milliseconds •  East coast to west coast calls typically experienced delays of 31.3-47.2 milliseconds
• Delays were due to carrier outages and reliability problems - they happened at least once a day for six out of seven carriers
• Overall application availability was 97%

Enterprise customers typically don't fare much better, experiencing 99.2% application availability, with blackouts occurring approximately 26% of the time (or 91minutes per month) and significantly reduced performance--brownouts--occurring 74% of the time (or 254 minutes per month).

Frustrating? Very. The public IP network represents a cheap channel for delivering voice calls and other voice-related applications. It promises to reduce capital equipment and operating costs, as well as dramatically lower long-distance charges. For these reasons, enterprise network managers are attacking poor availability with a number of tools designed to improve QoS and bandwidth utilization, including:

Quality of service scheduling techniques allow IT managers to give voice traffic priority. For example, DiffServ marking enables traffic to be classified as voice or data and then scheduled as priority traffic across the network. However, a strict prioritization approach tends to degrade under heavy loads. As the network architecture scales, QoS doesn't, and translating delay, jitter and loss requirements into specific classes of service is difficult. QoS techniques are suitable where bandwidth is constrained which is not the case in under-utilized core networks of many service providers.

Packet sequencing techniques create different types of circuits for different types of traffic and build itineraries for various flows admitted to the network. However, calls are only accepted if an itinerary satisfies each router's schedule; in addition, all network elements in the path must be capable of packet sequencing.

With packet shaping techniques, metering lights signal when congestion occurs and the packet shaping device attempts to re-shape flows once they begin to reach edge devices. While packet shaping can alleviate congestion at specific network points, such as the edge of a local network with limited-capacity public WAN connections, it can't avoid or fix problems that might occur elsewhere in the network.

Compression schemes simply stuff more data in each voice packet. They can help improve bandwidth efficiency, but they cannot help voice traffic avoid core network availability problems or outages unrelated to bandwidth issues.

Known as filtering techniques, buffer management schemes filter non-essential traffic, giving precedence to pre-defined traffic types. Buffer management techniques can help improve performance in networks with low-capacity links, but they provide little benefit across today's high-capacity service provider networks, and may even degrade performance.

Multiprotocol Label Switching (MPLS) helps service providers manage their public IP networks and provides multiple classes of service. However, from an enterprise perspective, MPLS links are simply public IP links using yet another technology. Early results from a RouteScience customer using both premium and best-effort services on an inter-city MPLS network suggest that MPLS-based public IP networks do not provide more than three nines of availability.

While many of these approaches may boost voice application performance over a LAN, WAN connections are the point of weakness. Either the entire path (including all links and networks) works, or the application fails. A company can't forestall problems in network fabric it doesn't own and no single service provider can deliver high enough quality for IP telephony traffic 24x7x365. Instead, the solution lies in looking beyond enterprise network boundaries with adaptive networking techniques. Adaptive networking takes advantage of the redundancy already implemented in all WANs. Just as load balancers distribute requests among servers in a data center, adaptive networking allows an enterprise to effectively use multiple WAN links or multiple classes of service on a single link. An adaptive networking approach considers application performance, determining the best path for IP telephony traffic to take and complementing existing QoS and other traffic management techniques.

Adaptive networking operates out of the direct traffic flow, monitoring end-to-end WAN performance by filtering NetFlow or span port data to identify IP telephony traffic, collect active endpoints, and categorize flows. It then uses a number of active and passive measurement techniques to determine the performance characteristics and uptime of each network path.

At the core of adaptive networking are specific application models that assess the relative impact of observed network impairments (such as latency, loss and jitter) and evaluate them in relation to each application type. These models incorporate the unique characteristics of Voice over IP, Web, enterprise, streaming media and real-time multimedia applications. This knowledge is then used to evaluate current network conditions and determine the best path through the WAN over which to steer traffic, based on each application's needs and link cost considerations.

Paths through the network are adjusted automatically in real time to meet business policies governing application availability, performance, minimum link commitment levels and cost. For example, policies can be designed to steer traffic to links that deliver the required performance at the lowest cost, once minimum commitment levels have been met. Adaptive networking solutions simultaneously document their assessment and decision-making processes in a variety of traffic and application-based reports, as well as provide live troubleshooting reports for those rare occasions when outages cannot be automatically bypassed. Each application model can also be fine-tuned to correspond to specific application environments.

While adaptive networking techniques will improve availability and performance for all business applications, they can deliver especially high return on investment for voice and multimedia applications. Adaptive networking delivers visibility into network paths and the data to back it up, allowing enterprises to confidently manage voice applications through all types of public IP networks, including the public Internet. With adaptive networking, enterprises can define policies that automatically manage--in real time--the balance between performance and cost.

 Service providers and enterprises that have already implemented adaptive networking have added a "nine" to the availability they gain on VoIP applications. Enterprises can combine public IP and private links to minimize costs and still improve performance and availability. For example, one financial services firm eliminated 90% of bad VoIP minutes it experienced. By optimizing links they already have, organizations have used adaptive networking techniques to eliminate network upgrades that were planned specifically to accommodate VoIP applications. And because adaptive networking continuously assesses all available paths and actively controls the network, users experience fewer brownouts and outages. For the IT team, these techniques enable staff to react quickly if necessary, maintain stability, and validate effectiveness of network controls.

A major online financial services firm collected end-to-end measurements for 11 days over its hybrid WAN platform that combined Frame Relay DS-3 links with Internet links with and without adaptive networking software. Voice traffic between the company's west-coast headquarters and an east-coast data center was monitored during the study. The firm experienced astounding results:

Network                     Bad Minutes              Availability

FR network                           5.4                         99.966%

Internet                              126.2                         99.203%

Optimized Internet                14.7                         99.907%

Optimized hybrid WAN          0.4                         99.997%

In another example, one of the world's largest hosting providers conducted a study using adaptive networking software to optimize its hosting environment. The results enabled the firm to safely consolidate multiple OC-3 (155 Mbps) links, save millions of dollars per month in bandwidth charges and evaluate traffic patterns, as well as validate hosting performance--all risk-free.

Public IP networks are still not ready to match the POTS network for dial-tone voice application availability. No matter. Adaptive networking techniques allow enterprises to extend their VoIP applications further than ever - even overseas - with acceptable quality and begin reaping the cost savings today. And they can simultaneously improve application performance and availability for all applications. There's no need to wait on the Internet. The route to toll-quality voice is now open.

Mansour Karam is the technical lead at RouteScience Technologies and can be reached at mans@routescience.com.

Visit RouteScience online.