Internet Traffic Dynamics

Propagation of Self-Similarity
"We demonstrate that if a TCP connection shares a bottleneck link with a self-similar background traffic flow, it propagates the correlation structure... This property is due to its congestion control algorithm, which adapts to fluctuations on several timescales."

The Internet is not just a collection of wires and routers; it is a complex adaptive system. One of the most striking features of internet traffic is its "fractal" nature—it looks bursty and self-similar across all timescales, from milliseconds to hours. But how does this pattern spread across the globe?

Our research focuses on the role of the TCP (Transmission Control Protocol), the fundamental language of the internet. We found that TCP is not merely a passive carrier of data; it actively propagates traffic patterns. Its congestion control algorithm acts as a "mirror," adapting to the fractal fluctuations of background traffic and transmitting them to other parts of the network.

The Mechanism of Propagation

We showed that TCP can be modeled as a linear system above a certain characteristic timescale (typically a few round-trip times). This means it faithfully "inherits" the statistical properties of the network environment.

Scaling analysis showing self-similarity in traffic
Fig 1. Evidence of Self-Similarity. Scaling analysis of traffic generated by a file transfer. The linearity of these plots on a log-log scale indicates a power-law relationship (Hurst parameter H ≈ 0.76), confirming the fractal nature of the traffic.
The Largest Exponent Rules:
A key finding is that if a connection traverses multiple congested links, the end-to-end traffic is characterized by the strongest self-similarity (the largest Hurst exponent) encountered along the path. Like a dominant gene, the most chaotic link defines the behavior of the entire connection.

Spreading Chaos

Furthermore, we discovered that self-similarity is contagious. A TCP stream adapting to a self-similar bottleneck can pass those statistical properties on to *other* TCP streams it interacts with later, even if those streams never touched the original bottleneck.

This explains why the "fractal" nature of internet traffic is so widespread. It is not just caused by user behavior (like file downloads); it is woven into the very fabric of the protocols that keep the internet running.

← Return to Main Page
Reference: Veres, A., Kenesi, Zs., Molnár, S. & Vattay, G. (2000). On the Propagation of Long-Range Dependence in the Internet. SIGCOMM '00, Stockholm, Sweden.