Abstract
Stateful serverless computing is a new paradigm within cloud computing. It successfully incorporates state management with serverless computing. Serverless computing is a form of cloud computing where the servers necessary for performing computation have been abstracted away, leaving the choice of where and how to perform a computation solely in the hands of the cloud provider. This abstraction simplifies the programming model for the cloud user, who can focus on business logic instead of scaffolding. It also offers the cloud provider greater freedom in how to manage the involved data centers, allowing for greater utilization of available resources.
In this thesis, I propose an autonomous platform for stateful serverless computing, provide a reference design and study the involved problems while providing their solutions. I focus on optimizing the entire system from the perspective of a cloud provider in terms of efficiency, cost and quality. The platform is able to autonomously adjust the supply of computing resources to meet fluctuations in demand without unnecessary waste. I show how to manage state in an efficient manner, which reduces latency while retaining flexibility in moving computations among servers. I further show how to manage a data cache in a cost-efficient manner, trading computation for storage. I present a new model for assigning computations to servers, allowing for higher utilization of available computing resources, thereby reducing the operational expenses of the cloud provider. I also show how to quickly solve this model, allowing for continuous redistribution of computations among servers to help maintain high resource utilization.
Merging theory and practice, I evaluate my designs both analytically and empirically. For empirical evaluation, I employ computational experiments, primarily through discrete-event simulation. While this work remains in its infancy, I believe that the presented concepts can be further refined into a working production system through dedicated, practical work. Some important questions remain unanswered, but hopefully they will one day be settled.
In this thesis, I propose an autonomous platform for stateful serverless computing, provide a reference design and study the involved problems while providing their solutions. I focus on optimizing the entire system from the perspective of a cloud provider in terms of efficiency, cost and quality. The platform is able to autonomously adjust the supply of computing resources to meet fluctuations in demand without unnecessary waste. I show how to manage state in an efficient manner, which reduces latency while retaining flexibility in moving computations among servers. I further show how to manage a data cache in a cost-efficient manner, trading computation for storage. I present a new model for assigning computations to servers, allowing for higher utilization of available computing resources, thereby reducing the operational expenses of the cloud provider. I also show how to quickly solve this model, allowing for continuous redistribution of computations among servers to help maintain high resource utilization.
Merging theory and practice, I evaluate my designs both analytically and empirically. For empirical evaluation, I employ computational experiments, primarily through discrete-event simulation. While this work remains in its infancy, I believe that the presented concepts can be further refined into a working production system through dedicated, practical work. Some important questions remain unanswered, but hopefully they will one day be settled.
| Original language | English |
|---|---|
| Supervisors/Advisors |
|
| Publisher | |
| Print ISBNs | 978-952-12-4053-9 |
| Electronic ISBNs | 978-952-12-4054-6 |
| Publication status | Published - 28 May 2021 |
| MoE publication type | G5 Doctoral dissertation (article) |