Abstract
In this paper, we address the problem of determining whether a bottleneck router on a given network path is using an AQM or a drop-tail scheme. We assume that we are given a source-to-sink path of interest—along which a bottleneck router exists—and data regarding the Round-Trip Times (RTT) and Congestion Window (CWND) sizes with respect to this flow. We develop a reliable classification algorithm that solely uses RTT and CWND information pertaining to a single flow to classify the queuing scheme, Tail Drop or AQM, used by the bottleneck router. We evaluate our method and present results that demonstrate our algorithm's highly accurate classification ability across a wide array of complex network topologies and configurations.
Introduction
Network congestion is typically managed at routers using queuing schemes. Traditional Tail Drop schemes wait for a buffer to fill before dropping incoming packets, often leading to "bufferbloat"—high latency caused by excessively large buffers. To combat this, Active Queue Management (AQM) schemes like PIE and RED were developed to preemptively drop packets, signaling the sender to slow down before the buffer overflows.
While AQM is generally superior, its actual deployment across the internet is largely unknown. Detecting whether a path uses AQM or Tail Drop is valuable for network tomography and for optimizing transport protocols. However, identifying these schemes from the "outside" using only end-to-end flow data is difficult due to network noise and varying topologies.
Methodology
The authors treat detection as a supervised classification problem, utilizing three main pillars:
- Randomized Topology Generation: To ensure generalization, a framework was built to generate thousands of diverse network topologies using a "path graph" backbone with random Erdős-Rényi graphs attached to each node.
- Simulation: These topologies were simulated using Mininet. For every topology, scenarios were run for both Tail Drop and PIE (a common AQM scheme) at the bottleneck router.
- Feature Engineering: Raw RTT and CWND data were smoothed using an Exponentially Weighted Moving Average (EWMA) filter. The researchers extracted 72 features, including first and second-order gradients, the number of local minima/maxima, and total variation ($L_2$-norm).
- Neural Network Classifier: After testing several models, a Neural Network with a single hidden layer of 14 neurons (using the LBFGS solver) was selected via randomized search.
Key Insights: RTT Signatures and Feature Importance
The research confirmed that Tail Drop creates a distinct "sawtooth" pattern in RTT (linear increases followed by sharp drops), while AQM results in shorter, more frequent bursts of fluctuations. Feature importance analysis revealed that RTT-based features are significantly more informative than CWND-based features. Specifically, statistics related to the gradients (the rate of change) of the time-series data were the most salient indicators of the underlying queuing scheme.
Results and Evaluation
The model was evaluated on two primary datasets:
- Training/Test Set: 1,100 unique network topologies (2,200 labeled data points).
- Complex Evaluation Set: Highly complex topologies with significantly more hosts, routers, and competing flows to test generalization.
On the standard test set, the classifier achieved 97% accuracy. When faced with the "Complex Topologies" dataset—which the model had never seen—it maintained a respectable 73% accuracy. This suggests the model learned the underlying physics of the queuing schemes rather than just memorizing specific network shapes.
Conclusion
This data-driven approach significantly outperforms previous heuristic-based methods, which typically achieved ~73% accuracy only on very simple, static topologies. While the current model primarily distinguishes between Tail Drop and PIE, future work will extend this to other schemes like RED and move toward multi-class classification. Further validation with real-world internet traffic will be necessary to account for non-TCP noise and diverse traffic patterns.
Summary updated