Edge Computing for IoT: Architecture Patterns That Work

A July 2026 paper presented at the IEEE International Conference on Edge Computing introduced EdgeFaaS — a function-based framework evaluated across a real testbed of 100+ geographically distributed IoT devices, edge servers, and cloud services. The research tested video analytics, federated learning, and audio classification workloads, and found that deployment decisions are highly workflow-specific: there is no single configuration that wins across all use cases. As IoT deployments scale into billions of connected devices, architects need principled frameworks for deciding which computation belongs at the device, the edge, and the cloud — and when to make those decisions dynamically.

Why Edge Computing Matters for IoT

Traditional cloud-centric IoT architectures work well at small scale: devices collect data, send it to the cloud, the cloud processes it and sends back results. But as IoT deployments grow — in factories, hospitals, smart cities, and agriculture — this model breaks down for several interconnected reasons:

• Latency: Cloud round-trips add 50–500ms of latency. For real-time control systems — industrial automation, autonomous vehicles, smart grid management — this is unacceptable.
• Bandwidth cost: Streaming raw sensor data from thousands of devices to the cloud is expensive and creates network bottlenecks at WAN boundaries.
• Privacy and data sovereignty: Many IoT deployments generate sensitive information (healthcare monitoring, factory floor data) that organizations are reluctant or legally prohibited from sending to third-party clouds.
• Reliability: Cloud-dependent systems fail when WAN connectivity drops. Edge-local processing enables graceful degradation with continued local operation.

Edge computing addresses these problems by placing computation closer to where data is generated — at the network edge, in local servers, gateways, or even on the devices themselves.

Image: File:IoT environmental monitoring system solution - Overview.jpg — Stephane Malhomme (CC BY-SA 4.0), via Wikimedia Commons

The Three-Tier IoT Architecture Model

The dominant reference architecture for IoT deployments follows a three-tier hierarchy that separates concerns by capability and proximity:

1. Device tier: Sensors, actuators, microcontrollers, and constrained IoT nodes that generate data and execute physical actions. Compute here is limited — typically ARM Cortex-M class processors with kilobytes of RAM.
2. Edge tier: Gateways, edge servers, and local compute clusters (often x86 or ARM Cortex-A class, running Linux) that aggregate data from many devices, run local inference and processing, and decide what to send upstream.
3. Cloud tier: Centralized cloud infrastructure handling long-term storage, ML model training, global analytics, and management control planes.

The key architectural decision in any IoT system is: which computation runs where? This is not a one-time decision — it's a continuous optimization problem as workloads evolve, data volumes grow, and hardware capabilities change at the edge.

Function-Based Frameworks: The EdgeFaaS Approach

The EdgeFaaS framework (Vadnere et al., 2026) represents a more principled approach to this placement problem. Rather than binding applications to specific infrastructure from the start, EdgeFaaS introduces two key abstractions:

• Function virtualization: Compute tasks are packaged as portable functions that can execute on IoT devices, edge servers, or cloud resources with a consistent interface. The framework handles differences in underlying hardware transparently.
• Storage virtualization: Data access is abstracted across distributed storage layers, so functions can read and write data without knowing whether it's cached at the edge or stored in the cloud.

In the IEEE EDGE 2026 evaluation, EdgeFaaS was used to deploy three representative IoT workflows: a video analytics pipeline (processing camera streams from distributed sensors), a hierarchical federated learning system (training ML models without sending raw data upstream), and an audio classification service. Across all three, the framework allowed engineers to vary function placement — IoT device vs. edge server vs. cloud — and directly observe the tradeoff between computation cost and communication bandwidth. The results demonstrated that optimal placement differs significantly by workload type, validating the case for flexible function-based architectures over rigid topology-bound designs.

Key Design Decisions When Building Edge IoT Architecture

Before choosing a framework or cloud provider edge offering, architects need to answer several foundational questions that shape all subsequent decisions:

1. What are your latency requirements? Classify workloads by latency tolerance: hard real-time (<10ms) must run on-device or on the immediate edge gateway. Soft real-time (10–100ms) can use a regional edge cluster. Best-effort workloads (>100ms) can tolerate cloud round-trips.

2. What data must stay local? Identify data that cannot leave a facility or jurisdiction due to privacy, regulatory, or contractual requirements. For healthcare IoT, patient vitals may need to stay on-premises. For factory IoT, proprietary production data may be restricted. Design your data flow with these constraints first, then optimize for performance within those boundaries.

3. How will you handle hardware heterogeneity? Edge environments are far messier than cloud environments. A single deployment may include Raspberry Pi nodes, NVIDIA Jetson devices, industrial PLCs, and x86 gateway servers. Container-based approaches with hardware abstraction layers reduce management burden, but you need to account for per-device capabilities when scheduling functions.

Security Considerations at the Edge

Edge IoT deployments dramatically expand the attack surface compared to centralized cloud architectures. Each edge node is a potential entry point, often in physically accessible locations. Essential security practices include:

• Device identity and attestation: Every edge node should have a unique hardware-rooted identity (TPM or secure element) enabling mutual TLS authentication before joining the mesh.
• Zero-trust between tiers: Don't assume that a message from the local network is trustworthy. Implement service-to-service authentication at every tier boundary.
• Encrypted communication: All inter-node communication — device to gateway, gateway to edge server, edge to cloud — should be encrypted in transit with modern cipher suites.
• OTA update capability: Plan for firmware and software patching of edge devices from day one. Retrofitting over-the-air update infrastructure to an existing deployment is significantly harder than designing it in from the start.

Frequently Asked Questions

What's the difference between edge computing and fog computing?

Fog computing (a term introduced by Cisco) refers specifically to a distributed computing layer between IoT devices and the cloud, often emphasizing hierarchical networks of gateways. Edge computing is a broader term that includes computation at or near data sources — from on-device inference to regional edge data centers. In practice, most industry conversations today use "edge computing" as the umbrella term, with "fog computing" appearing mainly in specific networking and industrial IoT contexts.

How does serverless (FaaS) fit into edge IoT architectures?

Function-as-a-Service platforms — including AWS Greengrass, Azure IoT Edge, and open-source options like OpenFaaS — extend the serverless model to edge hardware. The EdgeFaaS framework (2026) takes this further by providing consistent function virtualization across IoT, edge, and cloud tiers without requiring the same runtime environment everywhere. FaaS at the edge reduces operational complexity but introduces cold start latency — which matters more at the edge than in the cloud because real-time requirements are stricter. Keeping functions warm with minimum instance counts is the standard mitigation for latency-sensitive workloads.

Which cloud providers have the best edge IoT support in 2026?

AWS Greengrass, Azure IoT Edge, and Google Distributed Cloud Edge all offer mature managed services for edge IoT deployments. The best choice depends on your existing cloud commitments, geographic availability of edge nodes in your regions, and whether you need deep hardware-level integration (such as AWS Outposts for on-premises compute). Open-source alternatives like K3s (lightweight Kubernetes for edge), Eclipse Mosquitto (MQTT broker), and OpenFaaS provide more control but require greater operational investment from your team.

The Bottom Line

Edge computing for IoT is not a single technology choice — it's an architectural philosophy that places computation where it adds the most value given your latency, bandwidth, privacy, and cost constraints. We recommend starting with a clear classification of workloads by latency tolerance and data sensitivity, then designing an explicit three-tier architecture with deliberate data flow decisions at each tier boundary. Use function-based abstractions where possible — as the EdgeFaaS research demonstrates — because they make the compute-vs-communication tradeoff visible and adjustable as your deployment scales. Don't wait until you hit production scale to make these decisions: retrofitting centralized architectures for edge requirements is significantly harder than designing edge-aware systems from the start.

Disclaimer: This article is for informational purposes only. Technology landscapes change rapidly; verify information with official sources before making technical decisions.