Glasnost: Enabling End Users to Detect Traffic Differentiation

Marcel Dischinger, Massimiliano Marcon, Saikat Guha, P. Krishna Gummadi, Ratul Mahajan, Stefan SaroiuNSDI 2010 · OpenReview

In the heated debate over network neutrality, the missing link has often been transparency. While ISPs claim they only manage traffic to ensure network health, users often suspect their favorite apps—like BitTorrent or VoIP—are being intentionally slowed down. The challenge? Proving it.

The paper "Glasnost: Enabling End Users to Detect Traffic Differentiation" introduces a system designed to bridge this information gap, moving network measurement out of the lab and into the hands of everyday internet users.

TL;DR

Glasnost is a web-based measurement system that allows non-technical users to detect if their ISP is "differentiating" (throttling or blocking) specific application traffic. By emulating application-specific flows and comparing them against generic reference traffic, Glasnost provides a robust, low-barrier way to audit network neutrality at scale.

What Problem Does This Paper Address?

For years, traffic management policies were a "black box." ISPs might throttle BitTorrent to save bandwidth or deprioritize VoIP to protect their own telephony profits, but users had no way to distinguish between intentional throttling and general network congestion.

Existing research tools were often too complex for the average person, requiring software installation or administrative (root) privileges. The authors identified a need for a tool that is easy to use, robust against network noise, and evolvable to keep pace with changing ISP tactics.

Methods

Glasnost employs an active measurement strategy using a client-server architecture. When a user visits the Glasnost website, a small Java applet coordinates with measurement servers to run a series of tests.

Glasnost system architecture showing the interaction between the web server, measurement servers, and the client.

Paired Emulation

The core of the Glasnost system is paired emulation. The system sends two types of flows simultaneously:

  • Victim Flow: Emulates a specific protocol (e.g., BitTorrent) using actual protocol headers and payloads.
  • Reference Flow: Identical in size and timing to the victim flow but contains random bytes.
Comparison of BitTorrent and reference flows used in paired emulation.

By comparing the throughput of these two flows, Glasnost can identify protocol-specific differentiation. If the reference flow is significantly faster than the victim flow, it indicates that the ISP is likely using Deep Packet Inspection (DPI) to throttle specific applications.

Noise Filtering and Evolution

To avoid false accusations, Glasnost repeats tests and categorizes background noise. It discards results from paths that are too noisy to yield a reliable conclusion. Additionally, the authors developed a Trace-Emulate tool that takes a packet capture (pcap) of any application and automatically generates a new Glasnost test, allowing the community to quickly respond to new throttling targets.

Key Insights and Results

The study is backed by data from over 350,000 users across 5,800+ ISPs worldwide. Key findings include:

  • Prevalence of Throttling: Roughly 10% of users in the dataset experienced BitTorrent differentiation.
  • Asymmetric Throttling: ISPs are significantly more likely to throttle upstream (upload) traffic than downstream traffic, as P2P uploads consume more network resources.
  • The "User Impatience" Factor: Researchers found that completion rates significantly improved when tests were optimized to last roughly 6 minutes instead of 20.
  • Prioritizing False Positives: Glasnost was designed to be conservative, using a 20% throughput difference threshold to keep the false positive rate below 0.6%.

Limitations and Future Work

While successful, Glasnost faces challenges such as self-selection bias (users run it when they already suspect problems) and the potential for ISPs to whitelist Glasnost server IPs to hide throttling behavior. Furthermore, as ISPs move toward behavioral detection rather than simple protocol matching, simple emulation may need to evolve into more complex behavioral modeling.

Summary updated

BibTeX
@inproceedings{dischinger2010glasnost,
  author = {Marcel Dischinger and Massimiliano Marcon and Saikat Guha and P. Krishna Gummadi and Ratul Mahajan and Stefan Saroiu},
  title = {{Glasnost: Enabling End Users to Detect Traffic Differentiation}},
  booktitle = {NSDI},
  year = {2010},
  url = {https://openreview.net/forum?id=khEhMrn6Jr},
}
Figure