Year: 2024
Type: dissertation
Abstract: Large-scale cloud computing systems rely heavily on decision-making algorithms for critical system management tasks such as resource allocation, job scheduling, and power management.Manually crafted heuristics for these algorithms become increasingly untenable given the complexity and heterogeneity of modern cloud environments because the intricate interactions across diverse workloads, hardware platforms, and operating conditions make it exceedingly difficult to devise fixed heuristics that work well across all scenarios.Although machine learning techniques have been proposed to automatically learn optimized system management policies, existing approaches face practical limitations and lack the robustness necessary for production-grade cloud systems.This dissertation pioneers a novel abstraction-driven paradigm of efficient and robust online learning to fundamentally transform cloud systems management at scale.We develop a general framework that leverages deep reinforcement learning with system domain knowledge at its core to discover optimized management policies by continuously exploring and refining them through in situ interactions with cloud environments.To practically and robustly apply learning in cloud systems at scale, we further design two key abstractions: (1) A virtual agent abstraction that coordinates the distributed learned policies to resolve multi-agent interferences and stably converge on system-wide objectives, and (2) A meta learner abstraction that extracts generalizable policy embeddings that can rapidly adapt across the breadth of heterogeneous cloud applications and platforms.This abstraction-driven approach provides practical and extensible support for a wide range of online learning algorithms and diverse systems management tasks.Instantiated in systems like FIRM [1] (for microservices), SIMPPO [2] (for serverless computing), and µ-Serve [3] (for deep learning model serving), our innovative framework delivers order-of-magnitude improvements in resource efficiency, performance isolation, power optimization, and generalization compared to traditional heuristic-driven approaches.More profoundly, it establishes the foundations for practical and robust autonomous cloud systems management.Our contributions span the full stack, from mathematical models and optimizations to system design, implementation, and deployment.This incredible journey would not have been possible without the support and contributions of many cherished individuals.I am deeply grateful for the brilliant minds and warm hearts that made my PhD experience at the University of Illinois, Urbana-Champaign (UIUC) truly remarkable.First and foremost, I want to express my sincere gratitude to my PhD advisor, Prof.
Author: Qiu, Haoran
Institution: University of Illinois Urbana-Champaign
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7220815155
