The proliferation of data-intensive applications and the emergence of artificial intelligence (AI) workloads have driven a rapid increase in demand for computing resources. At the same time, increasing heterogeneity in hardware configurations and workload behaviors has made resource management increasingly challenging, as complex interdependencies between hardware and software can lead to congestion in shared resources. Such congestion degrades workload performance and increases energy consumption, forcing resource owners to leave hardware underutilized to mitigate the issue. This presents a significant challenge to the sustainability and efficiency of modern computing infrastructures.
This thesis addresses the challenges of resource management in heterogeneous cloud computing environments, with a particular focus on understanding and mitigating resource congestion and its effects on workload performance and cluster efficiency. It considers resource management at three levels of the cloud computing stack: the individual workload, the compute node, and the entire cluster. By investigating the interactions between workloads and hardware resources while considering the needs of different stakeholders, the thesis develops insights and approaches for improving resource utilization and cluster performance.
Across seven academic papers, this thesis progressively develops an understanding of resource congestion and its implications for cloud resource management. It first investigates the complex interdependencies between shared memory subsystem resources and workload performance. It then explores mechanisms for controlling these resources to mitigate the negative effects of congestion. Building on this foundation, the thesis introduces a novel observability framework that quantifies the impact of resource congestion on workload performance at runtime, providing the information needed to reason about resource behavior at the workload level. Finally, these insights are brought together in a congestion-aware and intent-driven approach to resource management that considers both the state of the underlying infrastructure and the needs of different stakeholders. The thesis concludes by discussing how these principles can inform resource management in increasingly heterogeneous cloud computing environments and the challenges that remain for the next generation of distributed computing systems.
Page Responsible: Frank Drewes 2026-11-22