Network Availability Based Service Differentiation

Mathilde Durvy, Christophe Diot, Nina Taft, Patrick ThiranIWQoS 2003 · DOI

TL;DR

This paper proposes a novel "degraded service" class for IP backbones that leverages underutilized bandwidth to offer a cheaper service without impacting existing "legacy" traffic. By introducing a new metric, "service availability," and simulating real-world failure patterns, they demonstrate that this approach can double network utilization and provide a clear service differentiation, especially in failure-prone environments.

What Problem Does This Paper Address?

Modern IP backbone networks are significantly over-provisioned to guarantee high performance, especially in the face of link failures. This leads to substantial unused bandwidth and low resource utilization, a major concern for network providers aiming to maximize return on investment. While Quality of Service (QoS) mechanisms exist, their complexity has hindered widespread commercial deployment. The authors aim to address this by proposing a simple, yet effective, QoS scheme that allows for service differentiation by introducing a "degraded" traffic class. This class would utilize the currently wasted bandwidth, offering a lower-cost service option to users, while ensuring that the critical "legacy" traffic (existing services) remains unaffected, even during network failures.

Methods

The core of the proposed solution is a simple two-class queuing system within routers: "Legacy Service: Fully Available" (FA) and "Degraded Service: Partially Available" (PA). FA traffic is given absolute priority, while PA traffic is only served when bandwidth is available after FA traffic has been accommodated. This is implemented using a strict priority scheduler with two drop-tail queues, requiring only a single bit marking in the IP header for traffic classification, thus avoiding complex signaling or admission control.

To evaluate this scheme, the authors employed a realistic simulation approach using the Sprint U.S. domestic backbone topology. They incorporated actual failure patterns observed in the Sprint network, distinguishing between partial and complete link failures. A key innovation is the introduction of "service availability" as a new SLA metric, defined as the fraction of time a service is available to a user, considering factors like packet loss and delay that could render applications unusable. They also used UDP traffic for simulations, arguing it provides a lower bound for TCP performance and that UDP loss rate is a convenient metric for network availability.

Datasets

The paper utilizes data from the Sprint U.S. domestic backbone network for:

  • Topology: The exact topology of the backbone network was used.
  • Failure Patterns: Real-world failure data from December 2001 to March 2002 was analyzed and used to construct failure scenarios for simulations.
  • Traffic Load: The paper mentions using gravity models to generate an FA traffic matrix that matches actual Sprint backbone link loads.

No new benchmark datasets were introduced; rather, real-world network characteristics were used for simulation.

Key Insights and Ideas

  • Leveraging Over-provisioning: The most significant insight is that the inherent over-provisioning in backbone networks can be actively utilized to offer a new service tier, rather than just being a costly redundancy.
  • Service Availability as a Key Differentiator: The introduction of "service availability" as an SLA metric is crucial. It moves beyond traditional metrics like packet loss and delay to capture the user's actual experience, especially in failure-prone environments.
  • Failure Patterns Drive Differentiation: Contrary to initial assumptions, link failures, particularly their frequency and duration, become the primary driver for differentiating between the FA and PA services. The PA service's performance is highly sensitive to network availability.
  • Simplicity is Key: The proposed QoS mechanism is remarkably simple, relying on basic priority queuing and IP header marking, making it practical for deployment in existing infrastructure.
  • PA Traffic Can Significantly Increase Utilization: The simulations show that adding PA traffic can at least double network utilization without negatively impacting FA traffic performance.

Key Results

  • The proposed two-class QoS scheme can be implemented with minimal changes to existing router hardware and protocols.
  • FA traffic performance remains stable and unaffected by the addition of PA traffic, even under high PA loads and network failures.
  • PA traffic performance (delay and loss) is sensitive to PA load and network availability. At low PA loads, network availability is the dominant factor; at high loads, PA load becomes more influential.
  • Service availability for PA traffic can drop significantly during network failures, highlighting its "degraded" nature. For instance, moving from 100% to 95% network availability can reduce PA service availability from 100% to below 50%.
  • With PA load equal to FA load, PA SLA is only marginally degraded, and service availability remains high (e.g., 98.5%).
  • Carriers can potentially quadruple the existing traffic load with a degraded service, depending on the network availability and the desired SLA for the PA class.

Limitations and Future Work

  • Simulation Simplifications: The authors acknowledge making simplifications, such as reducing link capacities to 10Mb/s for simulation tractability, though they argue relative performance is maintained. They also used UDP traffic, which provides a lower bound for TCP performance.
  • Probabilistic Failure Model: A probabilistic model for backbone failures is still needed for more accurate SLA estimations and a refined notion of network availability.
  • Intra-PoP Failures: The simulations did not consider intra-PoP failures, assuming their impact is minimal due to fully meshed core routers.
  • Service Availability Definition: The definition of service availability could be further refined to target specific applications that are likely to generate a large share of degraded traffic.
  • Future Work: The authors suggest developing statistical models for resource failures and refining the service availability metric. They also point to marketing studies as necessary to determine revenue potential from selling degraded services.

Summary updated

BibTeX
@inproceedings{durvy2003network,
  author = {Mathilde Durvy and Christophe Diot and Nina Taft and Patrick Thiran},
  title = {{Network Availability Based Service Differentiation}},
  booktitle = {IWQoS},
  year = {2003},
  doi = {10.1007/3-540-44884-5\_17},
}
Figure