Towards Elastic, High-Performance, Stateful Serverless Systems

Serverless computing is a transformative cloud computing paradigm. Serverless systems are automatically provisioned and elastically scaled by the cloud provider. The provider bills systems using a pay-as-you-go pricing model, which charges systems only for the resources they actually utilize, such as execution time and memory consumption, rather than for provisioned resources. While originally intended for stateless, lightweight tasks, serverless computing is increasingly being used to support more complex stateful systems. These systems introduce issues related to process coordination, state consistency, and performance variability that cannot be addressed using traditional serverless approaches. This dissertation investigates how to build efficient, cost-effective, elastically-scalable, performant, and stateful serverless systems. We describe the design, implementation, and evaluation of several stateful systems and frameworks. Through these efforts, we demonstrate that naively porting complex systems to a serverless architecture often results in high cost and poor performance. Instead, such systems require careful co-design of coordination strategies, state management techniques, and resource usage policies that are tailored to the constraints and strengths of the serverless model. Our contributions include: (1) λFS, a scalable serverless metadata service for distributed file systems that leverages the serverless elasticity of Functions-as-a-Service (FaaS) to provide fault-tolerant, parallel metadata operations; (2) NotebookOS, a serverless-computing enhanced Jupyter Notebooks-as-a-Service (NaaS) platform that addresses the inefficiencies of interactive deep learning training (IDLT) workloads in production notebook environments. NotebookOS introduces a replicated Raft-synchronized kernel design and a dynamic GPU time-aware scheduling mechanism to dramatically increase GPU utilization and improve interactivity; and (3) Sleipnir, a new FaaS runtime platform that integrates core mechanisms uncovered through this research, including fine-grained control over function state, metadata management, and execution scheduling, to directly support the construction of sophisticated, stateful FaaS systems. Together, these systems demonstrate the feasibility and benefits of building stateful systems on serverless infrastructure, while also offering new abstractions and techniques that advance the state of the art in serverless computing.

Author

• Carver, Benjamin Richard

• Cloud Computing
• Distributed File Systems
• Distributed Systems
• Functions-as-a-Service
• Machine Learning
• Serverless Computing

Identifier

hdl:1920/15142

OAI identifier oai:identifier

oai:MARS:1920/15142