1320 Edge Review: Architecture and Reference Model
1320.1 Learning Objectives
By the end of this chapter, you will be able to:
- Explain the Edge-Fog-Cloud Continuum: Describe how data flows through multiple processing tiers with increasing latency but decreasing bandwidth requirements
- Apply the Seven-Level IoT Reference Model: Map processing capabilities, latency characteristics, and use cases to each level
- Identify Processing Trade-offs: Evaluate latency, bandwidth, processing power, and cost at each architectural tier
- Design Tiered Architectures: Apply the golden rule of edge computing to determine optimal processing placement
1320.2 Prerequisites
Before studying this chapter, complete:
- Edge Compute Patterns - Core edge architectures
- Edge Data Acquisition - Data collection at the edge
- IoT Reference Models - Architecture foundations
Think of edge computing like a postal system:
- Edge (Level 1-3): Your local post office - handles urgent letters quickly, sorts mail before sending
- Fog (Level 4-5): Regional distribution center - accumulates mail from multiple local offices, makes routing decisions
- Cloud (Level 6-7): National headquarters - handles complex logistics, long-term records, cross-country coordination
Data flows the same way: urgent processing happens locally, while complex analysis travels to centralized systems.
1320.3 Edge-Fog-Cloud Architecture Overview
The following diagram illustrates the complete edge computing continuum, showing how data flows from sensors through multiple processing tiers to the cloud, with increasing latency but decreasing bandwidth requirements at each level.
1320.4 Seven-Level IoT Reference Model
The following table summarizes the seven-level IoT reference model, mapping each level to its processing capabilities, latency characteristics, and typical use cases.
| Level | Name | Processing Capabilities | Latency | Data Volume | Typical Use Cases |
|---|---|---|---|---|---|
| Level 1 | Physical Devices & Controllers | Raw sensing, basic actuation, signal conditioning | <1 ms | Very High (GB/hour) | Sensor sampling, emergency shutoffs, real-time control |
| Level 2 | Connectivity | Protocol translation, network routing, device addressing | <1 ms | Very High | Data transmission, network protocols, device communication |
| Level 3 | Edge Computing (Fog) | Evaluation (filtering), Formatting (standardization), Distillation (aggregation), Assessment (thresholds) | 1-10 ms | High (MB/hour) | Data reduction (100-1000x), downsampling, statistical aggregation, anomaly detection |
| Level 4 | Data Accumulation | Time-series storage, buffer management, data persistence | 10-100 ms | Medium | Local databases, recent data cache, query processing |
| Level 5 | Data Abstraction | Data modeling, semantic integration, format conversion | 10-100 ms | Medium | Data normalization, schema mapping, API abstraction |
| Level 6 | Application | Business logic, analytics, ML inference | 100-500 ms | Low (MB/day) | Dashboards, reporting, predictive models |
| Level 7 | Collaboration & Processes | Cross-system integration, workflow automation, enterprise services | 100-500 ms | Very Low | ERP integration, business processes, multi-tenant services |
1320.4.1 Key Data Reduction Example
Scenario: 500 vibration sensors, 1 kHz sampling, 16-byte readings
| Processing Stage | Data Rate | Reduction | Operations Applied |
|---|---|---|---|
| Raw Sensors (Level 1) | 28.8 GB/hour | Baseline | None |
| After Downsampling (Level 3) | 288 MB/hour | 100x | Frequency: 1 kHz to 10 Hz |
| After Aggregation (Level 3) | 2 MB/hour | 14,400x | Statistical summarization (100:1) |
Cost Impact: $25,000/year savings in cloud ingress costs (@$0.10/GB)
1320.4.2 Processing Trade-off Summary
| Factor | Edge Layer | Fog Layer | Cloud Layer |
|---|---|---|---|
| Latency | <1 ms (Best) | 10-100 ms (Moderate) | 100-500 ms (Highest) |
| Bandwidth | Very High (Worst) | Medium | Low (Best) |
| Processing Power | Limited | Moderate | Unlimited |
| Data Retention | Seconds-Minutes | Hours-Days | Unlimited |
| Cost per Node | Low | Medium | High (centralized) |
| Scalability | Distributed | Regional | Global |
| Use Cases | Real-time control, safety | Analytics, ML inference | Training, long-term storage |
1320.5 Architecture Design Principle
The Golden Rule of Edge Computing: Process data as close to the source as possible, but only as close as necessary.
- Edge (Level 1-3): Latency-critical operations, data reduction, real-time decisions
- Fog (Level 4-5): Regional analytics, ML inference, medium-term storage
- Cloud (Level 6-7): Deep analytics, model training, historical analysis, global coordination
1320.6 Knowledge Check: Architecture Concepts
Question: A factory deploys 500 vibration sensors sampling at 1 kHz (1000 Hz) with 16-byte readings. An edge gateway downsamples to 10 Hz and aggregates 100 sensors into summary statistics (200 bytes per aggregation). What is the data reduction from sensor to cloud per hour?
Explanation: Letโs calculate the data reduction through Level 3 edge processing:
Raw Sensor Data (Level 1):
- 500 sensors x 1000 Hz x 16 bytes = 8,000,000 bytes/second = 8 MB/s
- Per hour: 8 MB/s x 3600 seconds = 28,800 MB/hour = 28.8 GB/hour
After Downsampling (Level 3 - Gateway):
- Downsample 1000 Hz to 10 Hz (100x reduction in frequency)
- 500 sensors x 10 Hz x 16 bytes = 80,000 bytes/second = 80 KB/s
- Per hour: 80 KB/s x 3600 = 288,000 KB = 288 MB/hour
After Aggregation (Level 3 - Gateway):
- Aggregate 100 sensors into 1 summary record (200 bytes)
- Number of groups: 500 sensors / 100 = 5 groups
- 5 groups x 10 Hz x 200 bytes = approximately 2 MB/hour
Data Reduction: 28,800 MB/hour / 2 MB/hour = 14,400x reduction
This demonstrates Level 3 Edge Computing capabilities:
- Distillation/Reduction: Downsample high-frequency data (1 kHz to 10 Hz)
- Aggregation: Combine multiple sensor streams into statistical summaries
- Formatting: Standardize output format for cloud consumption
Question: An edge gateway receives data from 200 sensors. Level 3 processing applies evaluation (filtering 20% of bad data), formatting (standardizing), and distillation (aggregating 10 readings into 1 summary). The gateway buffer holds 100 readings. What happens when the 101st reading arrives before aggregation runs?
Explanation: This demonstrates Level 3 Edge Computing buffer management using FIFO (First In, First Out).
When the buffer reaches capacity (100 readings):
- New reading (101st) arrives
- Oldest reading is removed from the buffer
- New reading is appended to buffer
- Buffer size remains at 100
Why FIFO is appropriate for Level 3 Edge:
- Recency priority: Recent data is more relevant for real-time analytics
- Graceful degradation: System continues operating under high load
- No retries needed: Avoids network congestion from retransmissions
- Memory bounded: Prevents memory exhaustion on resource-constrained edge devices
Mitigation strategies for high-velocity data:
- Increase buffer size (1000 instead of 100)
- Faster aggregation cycles
- Multi-level buffering per sensor type
- Backpressure signaling to sensors
- Priority queueing for critical/anomaly data
1320.7 Common Misconception: Edge Equals Offline Processing
The Misconception: Many students believe edge computing means devices process data completely independently without cloud connectivity.
Reality - Hybrid Edge-Cloud Model:
Edge computing is about intelligent data reduction and latency-critical processing, NOT replacing cloud. The correct mental model:
- Edge processing: 95% of data volume (filtered locally)
- Cloud transmission: 5% of data (aggregated, important events)
- Cloud computation: 80% of ML training (requires historical data)
- Edge inference: 20% of ML (simple threshold-based decisions)
When to Use Each:
- Edge: Real-time safety shutdowns (<10ms), data reduction (100-1000x), privacy filtering
- Cloud: ML model training, historical analytics, cross-site correlation, firmware updates
- Wrong approach: Trying to do all ML training on edge devices, or sending all raw sensor data to cloud
Cost Impact of Misconception: Companies over-investing in edge infrastructure waste $50,000-$200,000 per deployment site, while companies under-utilizing edge spend $25,000-$80,000/year in unnecessary cloud ingress costs.
1320.8 Chapter Summary
The Edge-Fog-Cloud Continuum provides progressive data processing where latency increases (under 1ms at edge to 100-500ms at cloud) while bandwidth requirements decrease dramatically through data reduction at each tier.
The Seven-Level Reference Model guides processing decisions: Levels 1-2 handle physical sensing and connectivity, Level 3 performs edge computing (filtering, aggregation, format standardization), Levels 4-5 provide fog-layer storage and abstraction, and Levels 6-7 enable cloud analytics and enterprise integration.
Processing trade-offs must balance latency requirements, bandwidth constraints, processing power availability, data retention needs, and cost considerations when determining where to place computation.
The Golden Rule states: process data as close to the source as possible, but only as close as necessary, based on latency requirements and processing complexity.
1320.9 Whatโs Next
Continue to Edge Review: Data Reduction Calculations to learn how to calculate bandwidth savings and apply aggregation strategies for industrial IoT deployments.
Related chapters in this review series: