Date of Award

2019

Document Type

Open Access Dissertation

Degree Name

Doctor of Philosophy in Computer Science (PhD)

Administrative Home Department

Department of Computer Science

Advisor 1

Laura E. Brown

Advisor 2

Zhenlin Wang

Committee Member 1

Nilufer Onder

Committee Member 2

Allan A. Struthers

DOI

10.37099/mtu.dc.etdr/942

Abstract

Modern operating systems and microarchitectures provide a myriad of mechanisms for monitoring and affecting system operation and resource utilization at runtime. Dynamic runtime control of these mechanisms can tailor system operation to the characteristics and behavior of the current workload, resulting in improved performance. However, developing effective models for system control can be challenging. Existing methods often require extensive manual effort, computation time, and domain knowledge to identify relevant low-level performance metrics, relate low-level performance metrics and high-level control decisions to workload performance, and to evaluate the resulting control models.

This dissertation develops a general framework, based on the contextual bandit, for describing and learning effective models for runtime system control. Random profiling is used to characterize the relationship between workload behavior, system configuration, and performance. The framework is evaluated in the context of two applications of progressive complexity; first, the selection of paging modes (Shadow Paging, Hardware-Assisted Page) in the Xen virtual machine memory manager; second, the utilization of hardware memory prefetching for multi-core, multi-tenant workloads with cross-core contention for shared memory resources, such as the last-level cache and memory bandwidth. The resulting models for both applications are competitive in comparison to existing runtime control approaches. For paging mode selection, the resulting model provides equivalent performance to the state of the art while substantially reducing the computation requirements of profiling. For hardware memory prefetcher utilization, the resulting models are the first to provide dynamic control for hardware prefetchers using workload statistics. Finally, a correlation-based feature selection method is evaluated for identifying relevant low-level performance metrics related to hardware memory prefetching.

Share

COinS