Improving IoT Analytics through Selective Edge Execution

Apostolos Galanopoulos, Argyrios G. Tasiopoulos, George Iosifidis, Theodoros Salonidis, Douglas J. Leith

TL;DR

This paper introduces OnAlgo, an online, distributed algorithm that helps resource-constrained IoT devices decide when to "outsource" machine learning tasks to edge servers (cloudlets). By predicting the potential accuracy gain and balancing it against local power budgets and cloudlet capacity, the system achieves up to a 32% increase in accuracy and a 60% reduction in energy consumption compared to standard offloading strategies.

What Problem Does This Paper Address?

Modern IoT devices, such as wireless security cameras, are increasingly expected to perform complex analytics like image recognition. However, these devices face a "triple threat" of limitations:

  1. Compute Power: Local "low-precision" classifiers are often inaccurate.
  2. Energy: Transmitting data to a server drains precious battery life.
  3. Edge Capacity: While edge servers (cloudlets) are more powerful than IoT devices, they have limited capacity and cannot handle every single request from every device simultaneously.

Existing solutions often use simple heuristics—like offloading whenever a local model is "unsure"—or assume they have perfect knowledge of future network conditions. This paper tackles the challenge of making intelligent, real-time outsourcing decisions in unpredictable environments where data patterns and resource availability change constantly.

Methods

The authors propose a framework called OnAlgo, which treats the offloading decision as a constrained optimization problem.

  • The Predictor: Each IoT device runs a local "predictor" that estimates how much the accuracy would improve if a specific task were sent to the cloudlet. It doesn't just look at local confidence; it predicts the delta between local and edge performance.
  • Dual Ascent Optimization: To solve the problem without knowing future statistics (non-i.i.d. data), the authors use a Lagrange relaxation technique. This decomposes a massive system-wide problem into small, device-specific decisions.
  • Threshold Rule: The algorithm results in a simple, elegant threshold: a device offloads a task only if the predicted accuracy gain outweighs the "cost" (a combination of current energy drain and cloudlet congestion).
  • Running Averages: Because the system doesn't know the "true" average of network conditions or data arrivals, it uses running averages to adapt its decision threshold over time.

Datasets

The researchers evaluated their system using two classic computer vision benchmarks:

  • MNIST: A dataset of 60,000 grayscale images of handwritten digits (0-9).
  • CIFAR-10: A more complex dataset of 60,000 color images across 10 classes (animals, vehicles, etc.).
  • Classifiers: They tested these datasets using both K-Nearest Neighbors (KNN) and Convolutional Neural Networks (CNN) with varying depths (1 to 4 layers) to simulate different levels of local vs. edge capability.

Key Insights and Ideas

  • Selective is Better than Exhaustive: Blindly offloading every "low confidence" task (the ATO policy) actually hurts performance because it congests the cloudlet, causing tasks to be dropped and energy to be wasted.
  • Risk Aversion: The framework includes a "risk aversion" parameter ($\rho_n$), allowing users or designers to tune how aggressive the device should be in seeking cloudlet assistance.
  • Mathematical Robustness: A major contribution is the proof that OnAlgo achieves a "bounded and tunable optimality gap." In plain English: even though the algorithm makes decisions on the fly with imperfect info, it is mathematically guaranteed to get very close to the best possible theoretical performance over time.
  • The "Aha" Moment: The authors realized that the cloudlet's congestion level can be signaled back to devices via a single dual variable ($\xi$), effectively acting as a "price" for edge computation that naturally regulates traffic.

Key Results

The system was tested on a real-world wireless testbed using Raspberry Pis.

  • Accuracy Boost: OnAlgo increased classification accuracy by up to 32% compared to policies that don't account for resource constraints.
  • Energy Efficiency: It reduced energy consumption by up to 60% compared to "Accuracy-Threshold" policies.
  • Adaptability: In scenarios with many users, OnAlgo maintained higher accuracy because it "pruned" tasks that wouldn't benefit significantly from the edge, leaving room for tasks that actually needed the boost.
  • Convergence: The algorithm typically reaches a stable, near-optimal state in fewer than 300 iterations (time slots).

Limitations and Future Work

  • Predictor Dependency: The system's success relies heavily on the accuracy of the local "predictor." If the predictor cannot accurately guess the cloudlet's performance, the offloading decisions will suffer.
  • Static Model Assumptions: The paper assumes the ML models at the edge and local device are fixed. Future work could explore Dynamic Model Selection, where the system chooses between multiple edge models of varying complexity.
  • Bandwidth Constraints: While the model accounts for power and compute, it treats wireless links as high-capacity. In extremely bandwidth-constrained environments (like satellite IoT), the data transfer cost might need more granular modeling.

Summary updated

Abstract

A large number of emerging IoT applications rely on machine learning routines for analyzing data. Executing such tasks at the user devices improves response time and economizes network resources. However, due to power and computing limitations, the devices often cannot support such resource-intensive routines and fail to accurately execute the analytics. In this work, we propose to improve the performance of analytics by leveraging edge infrastructure. We devise an algorithm that enables the IoT devices to execute their routines locally; and then outsource them to cloudlet servers, only if they predict they will gain a significant performance improvement. It uses an approximate dual subgradient method, making minimal assumptions about the statistical properties of the system's parameters. Our analysis demonstrates that our proposed algorithm can intelligently leverage the cloudlet, adapting to the service requirements.
BibTeX
@inproceedings{galanopoulos2020improving,
  author = {Apostolos Galanopoulos and Argyrios G. Tasiopoulos and George Iosifidis and Theodoros Salonidis and Douglas J. Leith},
  title = {{Improving IoT Analytics through Selective Edge Execution}},
  booktitle = {ICC},
  year = {2020},
  url = {https://openreview.net/forum?id=SD2aIU8OalS},
}

Loading related content...

Figure