Bartendr: a practical approach to energy-aware cellular data scheduling

Aaron Schulman, Vishnu Navda, Ramachandran Ramjee, Neil Spring, Pralhad Deshpande, Calvin Grunewald, Kamal Jain, Venkata N. PadmanabhanMobiCom 2010 · DOI

TL;DR

Cellular radios are power-hungry, especially in weak signal areas. This paper introduces Bartendr, a system that intelligently schedules data transfers to coincide with strong signal periods, significantly reducing energy consumption for applications like email syncing and media streaming. By learning and predicting signal strength patterns along common routes, Bartendr achieves substantial energy savings without relying on power-intensive GPS.

What Problem Does This Paper Address?

Mobile devices are increasingly reliant on cellular data, but cellular radios are a major drain on battery life. This drain is exacerbated in areas with weak signals, where the radio must work harder (increasing power consumption) and transmits data at a slower rate, leading to a much higher energy cost per bit. While some applications, like real-time communication (e.g., VoIP), cannot tolerate delays, many others, such as email syncing, photo uploads, and media streaming, can be scheduled more flexibly. The challenge is to predict when strong signal conditions will occur without consuming excessive energy for the prediction itself, and then to effectively schedule these flexible data transfers.

Methods

Bartendr tackles this problem with a two-pronged approach:

  • Signal Strength Prediction using "Tracks": Instead of relying on energy-intensive GPS, Bartendr learns and predicts signal strength by creating "tracks." These tracks are essentially historical recordings of signal strength and base station IDs encountered along a specific route. By matching the current signal and base station information to these stored tracks, Bartendr can infer its approximate location on a known path and predict future signal strength. This approach leverages the observation that signal strength patterns along a path are often consistent, even across different days.
  • Energy-Aware Scheduling Algorithms:
    • Syncing: For applications like email sync, Bartendr uses a threshold-based approach. It predicts when the signal strength will exceed a certain threshold (e.g., -75 RSSI) and schedules the sync operation to occur during that predicted window. This allows the device to sleep for extended periods, waking up only when favorable signal conditions are anticipated.
  • Streaming: For streaming applications, Bartendr employs a dynamic programming algorithm. This algorithm considers predicted signal strength, corresponding throughput, and the "tail energy" cost (the energy consumed by the radio lingering in an active state after transmission) to find the optimal schedule for downloading data chunks. It aims to download data during periods of strong signal and high throughput, while also accounting for the cost of radio wake-ups and the need to meet playback deadlines.

Datasets

The paper utilizes data collected from actual driving experiments on four different cellular networks (Sprint, Reliance, Verizon, AT&T) across two metropolitan areas (Bangalore, India, and Washington D.C., USA). This data includes:

  • Signal strength measurements (both fine-grained RSSI and coarse "bars").
  • Instantaneous throughput measurements.
  • Power consumption measurements of mobile devices (Samsung SGH-i780, Palm Pre, Samsung Omnia).

The authors also created "signal tracks" from these experiments, which serve as the primary dataset for training and evaluating their prediction and scheduling algorithms.

Key Insights and Ideas

  • Signal strength is a critical factor in cellular energy consumption: The paper quantifies that weak signals can increase communication energy per bit by up to 6x compared to strong signals.
  • Location alone is insufficient for signal prediction: Hysteresis in cellular handoff mechanisms means that signal strength at a given location can vary depending on the direction of travel. This motivates the use of "tracks" that incorporate direction.
  • "Tracks" offer a practical and energy-efficient prediction mechanism: By learning signal patterns along common routes, Bartendr can predict future signal strength without the high energy cost of GPS.
  • Tail energy is a significant factor in cellular communication costs: The energy consumed by the radio lingering in an active state after a transmission needs to be accounted for in scheduling algorithms.
  • Flexible applications are key to energy savings: Applications that can tolerate some delay in communication are prime candidates for energy-aware scheduling.

Key Results

  • Syncing: Bartendr achieved energy savings of up to 10% for email sync operations, even when no actual data was downloaded. This highlights the savings gained solely from reducing radio power during strong signal periods.
  • Streaming: For streaming applications, Bartendr demonstrated significant energy savings of up to 60%. This is attributed to both reduced radio power and increased throughput during strong signal conditions, allowing for faster data downloads.
  • Prediction Accuracy: The "signal track" prediction method showed lower average prediction error than a simple predictor that assumes future signal will be the same as the current signal, especially for predictions beyond 20 seconds into the future.

Limitations and Future Work

  • Track Ambiguity: While generally effective, the "closest match" heuristic for locating a device on a track can be ambiguous in environments with sparse terrain or when signal strength doesn't change frequently.
  • Mobility Variability: Differences in current travel speed or stop durations compared to historical tracks can lead to prediction errors, though the system can recompute schedules dynamically.
  • Future Work:
    • Investigating more sophisticated signal strength prediction approaches, potentially incorporating accelerometer data.
  • Exploring alternate energy conservation models, such as leveraging always-on radio firmware or provider cooperation.
  • Developing better APIs for applications to access and utilize radio energy cost information.
  • Improving track aggregation and considering explicit modeling of mobility variations for more accurate predictions.
  • Exploring particle filter-based approaches for more robust position inference and prediction, especially in scenarios with conflicting information.

Summary updated

BibTeX
@inproceedings{schulman2010bartendr,
  author = {Aaron Schulman and Vishnu Navda and Ramachandran Ramjee and Neil Spring and Pralhad Deshpande and Calvin Grunewald and Kamal Jain and Venkata N. Padmanabhan},
  title = {{Bartendr: a practical approach to energy-aware cellular data scheduling}},
  booktitle = {MobiCom},
  year = {2010},
  doi = {10.1145/1859995.1860006},
}

Loading related content...

Figure