Zhanghao Wu

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2024-48

May 5, 2024

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-48.pdf

In an era where digital infrastructure increasingly relies on cloud computing, the need for flexible workload migration across clouds has become crucial. This need is particularly pressing with the recent surge in artificial intelligence (AI), which is impacted by global GPU shortages and geopolitical technology restrictions. The traditional cloud computing model, characterized by strong customer lock-in due to proprietary service interfaces and data gravity, limits the ability of businesses to adapt to these changes.

This dissertation extends and explores a recent rising concept, Sky Computing, as a transformative approach to mitigate these limitations. Sky Computing redefines the interaction between users and cloud services, proposing a unified "Sky of Computing" instead of isolated providers. This model leverages intercloud brokers to abstract the underlying cloud services, improving workload migration across clouds. We extensively explore the architectural and practical implementations of Sky Computing, leading to the development of an open-source intercloud broker, SkyPilot. SkyPilot demonstrates significant enhancements in optimizing and managing batch jobs, offering substantial cost savings, which is later extended to serving workloads, especially for AI. Further, this dissertation examines broker policy designed for deadline-sensitive jobs, implementing effective policies on SkyPilot that enable the utilization of unreliable but cost-effective spot instances while still meeting the deadline. Through real-world applications, Vicuna and SkyPilot Serving, we demonstrate how Sky Computing can support AI workloads, paving the way for further research.

Overall, we not only underscore the challenges faced by current cloud computing but also pioneer an adaptable approach through Sky Computing. This dissertation is an early step towards a more integrated and flexible cloud ecosystem, aligning technical innovation with market needs.

Advisors: Ion Stoica

---

BibTeX citation:

@phdthesis{Wu:EECS-2024-48,
    Author= {Wu, Zhanghao},
    Title= {Sky Computing with Intercloud Brokers},
    School= {EECS Department, University of California, Berkeley},
    Year= {2024},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-48.html},
    Number= {UCB/EECS-2024-48},
    Abstract= {In an era where digital infrastructure increasingly relies on cloud computing, the need for flexible workload migration across clouds has become crucial. This need is particularly pressing with the recent surge in artificial intelligence (AI), which is impacted by global GPU shortages and geopolitical technology restrictions. The traditional cloud computing model, characterized by strong customer lock-in due to proprietary service interfaces and data gravity, limits the ability of businesses to adapt to these changes.

This dissertation extends and explores a recent rising concept, Sky Computing, as a transformative approach to mitigate these limitations. Sky Computing redefines the interaction between users and cloud services, proposing a unified "Sky of Computing" instead of isolated providers. This model leverages intercloud brokers to abstract the underlying cloud services, improving workload migration across clouds. We extensively explore the architectural and practical implementations of Sky Computing, leading to the development of an open-source intercloud broker, SkyPilot. SkyPilot demonstrates significant enhancements in optimizing and managing batch jobs, offering substantial cost savings, which is later extended to serving workloads, especially for AI. Further, this dissertation examines broker policy designed for deadline-sensitive jobs, implementing effective policies on SkyPilot that enable the utilization of unreliable but cost-effective spot instances while still meeting the deadline. Through real-world applications, Vicuna and SkyPilot Serving, we demonstrate how Sky Computing can support AI workloads, paving the way for further research.

Overall, we not only underscore the challenges faced by current cloud computing but also pioneer an adaptable approach through Sky Computing. This dissertation is an early step towards a more integrated and flexible cloud ecosystem, aligning technical innovation with market needs.},
}

EndNote citation:

%0 Thesis
%A Wu, Zhanghao 
%T Sky Computing with Intercloud Brokers
%I EECS Department, University of California, Berkeley
%D 2024
%8 May 5
%@ UCB/EECS-2024-48
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-48.html
%F Wu:EECS-2024-48