Date of Award
8-2026
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Computer Science
Committee Chair/Advisor
Rong Ge
Committee Member
Thomas RW Scogland
Committee Member
Jacob Sorber
Committee Member
Zhenkai Zhang
Abstract
High performance computing (HPC) is dominated by heterogeneous systems that mainly derive performance from GPU accelerators. A majority of HPC applications have gravitated towards GPUs, which require explicit programming. Historically, programming explicitly to take advantage of GPUs has proven a significant barrier in fully utilizing GPU acceleration. Unified Memory (UM) is a technology designed to lower the barrier of entry to GPU programming by merging all memory domains of a system. While UM provides easier access to the capabilities of heterogeneous systems, the performance cost of UM greatly detracts from the benefit. Additionally, UM is implemented on a vendor-by-vendor basis and each implementation provides unique benefits and detractors. In order to quickly and fully utilize HPC systems, it is critical to understand UM design and mitigate the performance gap with explicit GPU programming.
In this dissertation, we explore the design and performance implications of Shared Virtual Memory (SVM), a leading UM implementation by AMD. We reveal the design of SVM and its interplay with the CPU side of UM memory management provided in Linux's Heterogeneous Memory Management (HMM) subsystem. We instrument low-level profiling to expose the performance impacts of SVM. We use our deep understanding of SVM’s implementation and performance analysis to make optimizations throughout SVM. We finally look at the next generation of GPU systems that physically unify memory between CPU and GPU processors in a single APU chip. As such, this work provides critical insight to improving the accessibility and performance for both current and future HPC systems.
Recommended Citation
Cooper, Bennett, "A Comprehensive Study and Performance Optimization of Shared Virtual Memory" (2026). All Dissertations. 4344.
https://open.clemson.edu/all_dissertations/4344
Author ORCID Identifier
0009-0003-0375-9418