Visibility

Megan Wachs, Jung‐Il Choi, Jungwoo Lee, Kannan Srinivasan, Zhe Chen, Mayank Jain, Philip Levis2007 · DOI

TL;DR

This paper introduces "visibility" as a crucial, yet often overlooked, design goal for sensornet protocols. They propose a quantitative metric to measure visibility (the energy cost of diagnosing failures) and demonstrate how designing for visibility can lead to simpler debugging, improved efficiency, and even better performance, as shown with their novel Pull Collection Protocol (PCP).

What Problem Does This Paper Address?

Sensornet deployments frequently fail to meet performance expectations, and a major hurdle is the difficulty in diagnosing why they fail. Existing debugging tools often require significant energy expenditure to extract information from the network. The authors argue that the responsibility for providing this crucial insight shouldn't solely fall on external tools, but rather on the network protocols themselves. They aim to shift the focus from purely performance metrics (throughput, latency, fairness) to include "visibility" – the ease with which network behavior and failures can be understood.

Methods

The core contribution is the introduction of a quantitative visibility metric. This metric is defined as the "energy cost of diagnosing the cause of a behavior in a protocol." They propose a decision-tree-based approach, similar to Sympathy, where the cost of diagnosing a failure is the sum of the energy required to answer diagnostic questions, weighted by the probability of needing to answer that question.

To demonstrate their concept, they:

  • Designed Pull Collection Protocol (PCP): A novel multi-hop collection protocol that aims to eliminate common causes of packet loss (like ingress drops) by using a "pull" mechanism instead of traditional "push." This design inherently simplifies failure diagnosis.
  • Optimized Deluge: They modified the Deluge dissemination protocol to improve its visibility, creating "V-Deluge." This involved identifying and eliminating non-local queries required for diagnosis.
  • Developed a Visibility Calculation Methodology: This involves:
  • Using wired back-channels in testbed experiments to log network events and identify failure causes.
  • Analyzing this data to estimate the probability distribution of different failure causes.
  • Constructing decision trees for diagnosis and calculating the visibility cost using their proposed metric.

Datasets

The paper does not explicitly mention or introduce new datasets. Instead, it relies on data collected from experiments conducted on the Motelab testbed (a community testbed for wireless sensor networks). These experiments involved running protocols like MultihopLQI, PCP, and Deluge/V-Deluge on networks of 31 to 61 motes. The "datasets" are essentially the logs and measurements generated during these controlled experiments.

Key Insights and Ideas

  • Visibility as a First-Class Design Goal: The paper strongly advocates for treating visibility not as an afterthought or solely the domain of debugging tools, but as a fundamental design principle for sensornet protocols.
  • Protocols Should Be Self-Diagnosing: The authors argue that protocols should be designed to inherently provide information about their own state and behavior, reducing reliance on external, energy-intensive diagnostic tools.
  • Visibility Can Improve Performance: Counter-intuitively, designing for visibility by eliminating failure causes can lead to better throughput and reliability, as demonstrated by PCP. This is because failures themselves negatively impact performance.
  • The Cost of Diagnosis Matters: The proposed visibility metric quantifies the energy cost of understanding a failure, providing a concrete way to compare protocol designs.
  • PCP's Pull Mechanism Simplifies Diagnosis: By eliminating ingress drops and other common failure modes, PCP's design makes it significantly easier to diagnose why packets are lost.

Key Results

  • PCP's Visibility: In a controlled testbed environment, PCP achieved a visibility cost of 0C (where C is the cost of an external query), compared to MultihopLQI's visibility cost of 1.615C. This highlights PCP's inherent diagnostic simplicity.
  • PCP's Performance: PCP achieved better goodput and end-to-end delivery reliability than MultihopLQI. For example, at comparable packet generation rates, PCP lost only 0.6% of packets, while MultihopLQI lost 9.6%.
  • V-Deluge's Efficiency: The visibility-enhanced Deluge (V-Deluge) disseminated a binary 29.8% faster and sent 45.6% fewer packets than the original Deluge, demonstrating that improved visibility can lead to increased efficiency.
  • Fairness Trade-offs: While PCP offers excellent visibility and reliability, its fairness was slightly lower than MultihopLQI at lower sending rates due to approximated subtree sizes.

Limitations and Future Work

  • Probability Distribution Estimation: The accuracy of the visibility metric relies on knowing the probability distribution of failure causes. The paper acknowledges that obtaining this distribution in real-world, dynamic environments is challenging and suggests adaptive measures to monitor and update this distribution.
  • External Dependencies: The current visibility metric assumes questions are answered by sending queries. Future work could explore piggybacking information on data packets to reduce query costs.
  • Protocol Isolation: The paper notes that sensornet systems often involve multiple protocols, and failures in one can affect others. Future work could focus on mechanisms to ensure better isolation between protocols for clearer diagnosis.
  • Dynamic Decision Trees: The optimal decision tree for diagnosis can change as failure distributions evolve. Future research could explore dynamically adapting these trees.
  • PCP's Visibility Nuances: While PCP is highly visible for packet loss, diagnosing why a node never sends packets can still be complex. The authors acknowledge this and suggest further work on handling such scenarios.

Summary updated

Abstract

This paper proposes a new sensornet protocol design goal: visibility. Visibility into behaviors at the network level will simplify debugging and ease the development process. We argue that increasing visibility is the responsibility of the network protocols themselves, and not solely the responsibility of existing debugging tools. We describe a quantitative visibility metric to evaluate and compare protocols, where visibility is defined as the energy cost of diagnosing the cause of a behavior in a protocol. The design and evaluation of Pull Collection Protocol, a novel multi-hop collection protocol, is an example of how to design for visibility without sacrificing throughput or node-level fairness. We also describe our optimizations for an existing protocol, Deluge, to increase its visibility and efficiency.
BibTeX
@misc{wachs2007visibility,
  author = {Megan Wachs and Jung‐Il Choi and Jungwoo Lee and Kannan Srinivasan and Zhe Chen and Mayank Jain and Philip Levis},
  title = {{Visibility}},
  year = {2007},
  doi = {10.1145/1322263.1322271},
}

Loading related content...

Figure