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.
Introduction
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: compute power, energy constraints, and limited edge capacity. While edge servers (cloudlets) are more powerful than IoT devices, they cannot handle every single request from every device simultaneously without becoming congested.
This paper introduces OnAlgo, an online, distributed algorithm that helps resource-constrained IoT devices decide when to "outsource" machine learning tasks to edge servers. By predicting the potential accuracy gain and balancing it against local power budgets and cloudlet capacity, the system achieves significant improvements in both accuracy and energy efficiency.
System Architecture and Methodology
The authors propose a framework where the offloading decision is treated as a constrained optimization problem. The architecture, as illustrated in Figure 1, involves several key components within each IoT device that facilitate intelligent decision-making.
- The Predictor: Each IoT device runs a local "predictor" that estimates the accuracy improvement (the delta) if a specific task were sent to the cloudlet instead of being processed locally.
- 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 threshold: a device offloads a task only if the predicted accuracy gain outweighs the "cost"—a combination of current energy drain and cloudlet congestion.
- Congestion Signaling: The cloudlet's congestion level is signaled back to devices via a single dual variable ($\xi$), acting as a "price" for edge computation that naturally regulates traffic.
Evaluation and Key Results
The system was evaluated using classic computer vision benchmarks, MNIST and CIFAR-10, with various classifiers (KNN and CNN) of varying depths to simulate different local vs. edge capabilities. Testing was conducted on a real-world wireless testbed using Raspberry Pis.
As shown in Figure 3, OnAlgo significantly outperforms standard offloading strategies like Accuracy-Threshold (ATO) and Random-Choice (RCO):
- Accuracy Boost: OnAlgo increased classification accuracy by up to 32% compared to policies that do not account for resource constraints.
- Energy Efficiency: It reduced energy consumption by up to 60% compared to ATO policies.
- Scalability: In scenarios with many users, OnAlgo maintained higher accuracy because it "pruned" tasks that would not benefit significantly from the edge, leaving room for tasks that actually needed the boost.
Discussion and Future Work
A major contribution of this work is the mathematical proof that OnAlgo achieves a "bounded and tunable optimality gap," meaning it is guaranteed to perform near-optimally even in unpredictable environments. The framework also includes a "risk aversion" parameter ($\rho_n$), allowing for tuning of how aggressively a device seeks cloudlet assistance.
However, the system's success relies heavily on the accuracy of the local predictor. Future work could explore dynamic model selection at the edge and more granular modeling of bandwidth constraints in extremely limited environments, such as satellite IoT.
Summary updated