Cloud, Fog, or Edge: Where Should You Compute for AI?

As the digital landscape evolves, the question of where to execute computational tasks—whether on the cloud, at the edge, or within fog computing environments—becomes increasingly critical. In our recent study, we explored the performance and efficiency of these computing paradigms across various scenarios to help guide optimal decision-making.

Video Encoding: The Power of the Edge

Video encoding is a resource-intensive process that benefits significantly from low latency and high computational efficiency. Our research found that edge devices, particularly the latest generation of single-board computers like the Raspberry Pi 4 and Jetson Nano, excel in video-on-demand encoding. These devices reduce raw video transfer times and perform encoding tasks more efficiently compared to older models and some cloud instances.

For continuous live stream encoding, cloud resources prove advantageous due to their lower encoding times, despite the potential for higher raw video transfer times. Cloud instances such as AWS m5a.xlarge offer the computational power needed for seamless live streaming.

Recommendation: Utilize edge devices for video-on-demand services to minimize transfer times. For live streaming, cloud instances are preferable to ensure low encoding times and smooth performance.

AI: Cloud Dominance and Edge Potential

Machine learning tasks vary widely in complexity and data requirements. Our evaluation used TensorFlow to train neural networks, revealing distinct advantages for different computing environments based on the scenario.

For large datasets and complex models with multiple layers, cloud resources are ideal. Their superior computational power ensures efficient training times and higher accuracy. In contrast, edge devices can handle smaller datasets and simpler models, offering a feasible alternative when low latency and local processing are priorities. Interestingly, the Jetson Nano stands out among edge devices, showing faster training times for convolutional neural networks compared to older single-board computers like the Raspberry Pi 3B.

Recommendation: For complex and large-scale machine learning tasks, offload computations to the cloud. Use edge devices for simpler models and smaller datasets to leverage local processing advantages.

In-Memory Data Analytics: Balancing Act

In-memory data analytics, critical for real-time data processing, benefits from the scalability and power of cloud and fog environments. Using Apache Spark, our study demonstrated that these environments offer a balanced approach, providing necessary computational power while maintaining manageable latency.

Recommendation: Employ cloud and fog environments for large-scale data analytics to achieve the best balance between performance and latency.

Conclusion

Choosing the right computational environment—cloud, fog, or edge—depends on the specific requirements of your task. Our research provides a roadmap for making informed decisions:

- Edge: Best for video-on-demand encoding and simpler machine learning models.

- Cloud: Optimal for live video streaming and complex machine learning tasks.

- Fog: Ideal for scalable in-memory data analytics.

By strategically leveraging these environments, you can enhance performance, reduce latency, and optimize resource utilization for your computational tasks.

For an in-depth look at our findings, you can access the full paper: https://ieeexplore.ieee.org/abstract/document/9321525/.