Year: 2024
Type: dissertation
Abstract: Machine learning (ML) training and inference systems encounter constraints in current computation environments due to increased ML model sizes, the fast-growing popularity of ML/AI, etc.In this thesis, we show how machine learning training and inference systems can be executed successfully and efficiently in constrained computation environments, such as limited-memory GPUs, on-premises clusters, and serverless environments, by using a novel combination of algorithms, optimizations, and well-reasoned system designs.Concretely, we propose (i) a system that enables large ML model training over multiple memory-constrained GPU devices via algorithms and system designs that achieve fast placements with a quality comparable to expert-designed placements, (ii) a system that enables efficient resource sharing among ML inference jobs in fixed-size on-premises clusters by making close-to-optimal autoscaling decisions quickly via several relaxation methods in optimization and prediction, and (iii) a system that enables cost-efficient distributed GNN training on constrained serverless execution environments by auto-tuning configuration via analytic model-based offline optimization and gray-box heuristic-based online optimization.I would like to express my sincere gratitude to my advisor, Prof. Indranil Gupta, for his support and guidance.I have enjoyed countless intellectual discussions with him and learned a lot about developing ideas, conveying them crisply and professionally, and many other aspects.I also appreciated his valuable advice on research directions, career paths after the PhD program, and even life.
Author: Jeon, Beomyeol
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: National Science Foundation
OpenAlex ID: W7220783224