Sky computing

Sky computing is a paradigm that aims to develop the cloud computing model further. It aims to combine existing clouds of different service providers into a comprehensive, interoperable Sky. The concept behind sky computing is to create a cloud of clouds that behaves in a similar way to the internet, which consists of a network of networks.^[1]

Description

Sky computing aims to achieve a complete abstraction of cloud resources from different providers so that applications and users can access these resources without having to worry about where the resources or services are located in the individual clouds. The key features of Sky Computing include:

• Cloud of clouds: a unified, interoperable cloud made up of numerous individual clouds.
• Levels of abstraction: These ensure the interoperability of clouds.
• Distributed infrastructure: A comprehensive infrastructure for cloud services.
• Dynamic scalability: Resources can be scaled dynamically across multiple clouds.
• Universality: Applications can be run in any cloud.

Typical technical components serve to standardize interfaces, enable the dynamic deployment of cloud services, and support ___location- and cost-aware data distribution^[2]:

Compatibility layer: This layer abstracts the services of different cloud computing providers and provides unified application programming interfaces (APIs). Open-source technologies such as Kubernetes, Docker, or Apache Spark are often used for this purpose. The goal is to reduce dependence on proprietary cloud services and enable portable, vendor-independent architectures.

Intercloud layer: Building on the compatibility layer, this layer enables the dynamic, policy-based deployment of cloud services across multiple providers. Users can usually specify criteria such as cost, ___location requirements (e.g., GDPR compliance), availability, or execution time. Cloud services are maintained in a central catalog under a unified naming convention and described with metadata such as provider, region, and price. This information makes it possible to take individual preferences into account when selecting and allocating services. In many cases, the intercloud layer includes a central billing system that handles payment processing without requiring end users to enter into separate contracts with each provider.

Peering mechanism: For data-intensive applications, it may be necessary to transfer large volumes of data between different cloud providers. Since downloading data (egress) is generally more expensive than uploading data (ingress), a peering mechanism takes these cost differences into account during data processing. It coordinates data distribution to suitable regions or cloud infrastructures, provided this complies with user specifications.

History

The concept of sky computing emerged from research in distributed systems and cloud computing in the late 2000s.^[3][1] The term was first coined in a 2009 publication.^[4]

The concept was further developed by Ion Stoica and Scott Shenker of UC Berkeley in 2021.^[5] The concept of the intercloud broker is being further developed at the Sky Computing Lab at UC Berkeley.^[6] For example, the open-source library SkyPilot has emerged from this work.^[7] Companies such as OpenRouter and Cortecs have built intercloud layers to distribute large language model (LLM) workloads across cloud providers.^[8][9]

Challenges

Although Sky Computing has the potential to transform the cloud computing landscape, several challenges remain. These include technical obstacles, data protection and security concerns, as well as regulatory issues. There is also a risk that an effective market may not develop, for example due to incomplete or missing information on services and pricing, or because of aggressive pricing strategies by dominant cloud providers.^[10]

References

1. Armbrust, Michael; Fox, Armando; Griffith, Rean; et al. (February 10, 2009). Above the Clouds: A Berkeley View of Cloud Computing (PDF) (Technical report). EECS Department, University of California, Berkeley. UCB/EECS-2009-28. Retrieved 2025-08-28.
2. Ion Stoica; Scott Shenker (2021). "From Cloud Computing to Sky Computing". Proceedings of the Workshop on Hot Topics in Operating Systems (HotOS '21). Ann Arbor, Michigan: Association for Computing Machinery. pp. 26–32. doi:10.1145/3458336.3465301.
3. Buyya, Rajkumar; Yeo, Chee Shin; Venugopal, Srikumar (2008). "Market-Oriented Cloud Computing: Vision, Hype, and Reality for Delivering IT Services as Computing Utilities". 2008 10th IEEE International Conference on High Performance Computing and Communications. pp. 5–13. arXiv:0808.3558. doi:10.1109/HPCC.2008.172. ISBN 978-0-7695-3352-0.
4. Keahey, Katarzyna; Tsugawa, Mauricio; Matsunaga, Andrea; Fortes, Jose (2009). "Sky Computing". IEEE Internet Computing. 13: 43–51. doi:10.1109/MIC.2009.94.
5. Wang, Stephanie; Hindman, Benjamin; Stoica, Ion (2021). "In reference to RPC: it's time to add distributed memory". Proceedings of the Workshop on Hot Topics in Operating Systems. HotOS '21. Ann Arbor, Michigan: Association for Computing Machinery. pp. 191–198. doi:10.1145/3458336.3465302. ISBN 978-1-4503-8438-4. Retrieved 2023-07-10.
6. "UC Berkeley Sky Computing – UC Berkeley Computer Science Dept". Retrieved 2023-07-10.
7. Sky Computing Lab. "SkyPilot – Run LLMs, Serve Models, and Run Batch Jobs on Any Cloud". GitHub. Retrieved 2025-05-22.
8. "OpenRouter – The Unified Interface for LLMs". Retrieved 2025-08-28.
9. "cortecs.ai – Sky Inference". Retrieved 2025-05-22.
10. Chasins, Sarah; Cheung, Alvin; Crooks, Natacha; et al. (2022). "The Sky Above The Clouds". arXiv:2205.07147.