Towards Describing Multi-fractality of Traffic Using Local Hurst Function

Ming Li, Sungbin Lim, Bai-Jiong Hu, Huamin FengLecture Notes in Computer Science 2007 · DOI · OpenAlex

Introduction

Network traffic characterization has long relied on the concepts of self-similarity and long-range dependence (LRD). Traditional models, such as Fractional Brownian Motion (fBm), utilize a constant Hurst index ($H$) to describe these properties. However, this "monofractal" approach assumes that the regularity or "smoothness" of traffic remains uniform over time. In practice, internet traffic is highly volatile and exhibits multifractal behavior, where scaling properties shift dynamically. This paper proposes a transition from static models to a multifractional framework using a time-varying local Hurst function, $H(t)$, to capture granular, point-by-point scaling behavior.

Methodology: From fBm to mBm

The core of this research involves bridging the gap between monofractal theory and multifractal reality through Multifractional Brownian Motion (mBm). Unlike fBm, which uses a constant $H$, mBm introduces a local Hurst function $H(t)$, also known as the local Hölder exponent.

The authors developed a numerical method to estimate $H(t)$ at specific points in time by analyzing the local growth of the traffic's increment process. This mathematical framework includes derivations for covariance and variance structures, specifically focusing on a "causal" (one-sided) version of the model. This causal approach is critical for real-time network systems where only historical and current data are available for prediction.

Experimental Analysis

The proposed $H(t)$ estimator was evaluated using standard real-world traffic traces (DEC-PKT). These traces represent typical TCP packet traffic in computer networks. By applying the estimator to these datasets, the researchers visualized the evolution of burstiness over thousands of packets.

Comparison of traffic bytes x(i) and the estimated local Hurst function H(t) for a DEC-PKT trace.

Key Insights and Results

The analysis yielded several significant findings regarding the nature of network traffic:

  1. Erratic Nature of $H(t)$: The local Hurst function is not a smooth curve; rather, it behaves like a random process itself. This suggests that the underlying nature of network traffic is constantly reinventing its scaling properties.
  2. Dynamic Burstiness: Since $H$ characterizes burstiness, a fluctuating $H(t)$ explains why network buffers may overflow unexpectedly even when average traffic levels appear manageable. The "texture" of the traffic changes in ways that static models cannot predict.
  3. Local Stationarity: While mBm is not globally stationary, its increment processes are locally stationary. This property allows for more robust mathematical modeling of short-term traffic behavior.
  4. Network Performance Implications: The fluctuations in $H(t)$ are directly linked to practical issues such as packet delay and congestion. The inherent randomness of the local Hurst function is a primary reason why network delay remains difficult to predict accurately.
Traffic bytes and local Hurst function for a different trace, illustrating the dynamic nature of burstiness.

As shown in the figures, $H(t)$ typically fluctuates significantly (e.g., between 0.7 and 0.85 for the DEC traces). These results confirm that a single, average $H$ value is an oversimplification that misses critical local behavior.

Conclusion and Future Directions

This work provides strong evidence for the multifractal nature of internet traffic and demonstrates the utility of the local Hurst function in describing it. While the researchers can now calculate $H(t)$ numerically, the development of an analytic model—a mathematical formula describing the evolution of $H(t)$ itself—remains an open challenge.

Future research will focus on finding statistical or deterministic models for $H(t)$ to enable the simulation of more realistic synthetic traffic. Additionally, the local Hurst function shows promise as a tool for pattern recognition, potentially offering a "signature" for identifying network anomalies such as DDoS attacks.

Summary updated

BibTeX
@inproceedings{li2007towards,
  author = {Ming Li and Sungbin Lim and Bai-Jiong Hu and Huamin Feng},
  title = {{Towards Describing Multi-fractality of Traffic Using Local Hurst Function}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2007},
  doi = {10.1007/978-3-540-72586-2\_143},
}
Figure