Abstract
dc:description.abstractModern computing platforms, from embedded systems in safety-critical environments to large-scale cloud servers, rely on shared memory hierarchies that introduce unpredictable interference among co-running applications. Such interference undermines performance guarantees for performance-critical and latency-sensitive workloads and often forces overprovisioning of critical hardware resources. While numerous studies have focused on providing temporal guarantees to critical applications by isolating workloads through resource partitioning, the use of fine-grained profiling to capture real application interactions with the underlying hardware has remained under explored in both domains, albeit for different reasons. This thesis demonstrates how profiling, analyzing, and managing based on fine-grained hardware counter data can enable transparent and effective control over shared hardware resources. The research spans two domains with distinct but related challenges: real-time embedded platforms and multi-tenant cloud servers. In the embedded context, a fine-grained profiling and control framework, E-WarP, is developed to capture an application’s time-varying memory bandwidth demands using hardware performance counters. The resulting profiles are used to construct cumulative memory envelopes that yield tighter worst-case execution time (WCET) bounds than conventional worst-case models. These insights are then used to enforce memory bandwidth quotas across CPUs and accelerators through a combination of software and hardware mechanisms, achieving observed execution times close to the predicted WCET. In the cloud domain, the same profiling principles are extended to highly dynamic, multi-core server environments through the ProPer framework. A low-overhead profiler captures both cache and memory behavior using widely available mechanisms. By recording hardware counters and function transition points across multiple runs, the profiler incrementally builds a detailed profile of an application’s interaction with the underlying hardware. Using this accumulated data, the framework automatically identifies phase-specific resource usage patterns and derives corresponding phase-aware allocation policies. These policies are enforced at runtime using Intel’s Cache Allocation Technology (CAT) and a custom user-space bandwidth regulator (USBW) that provides finer control than existing hardware mechanisms. Together, these efforts demonstrate that measurement-based, profile-driven resource management can bridge the gap between real-time determinism and cloud-scale efficiency. By systematically leveraging fine-grained performance data to inform allocation decisions, this work advances the state of the art in predictable and efficient resource sharing at the shared memory hierarchy. The resulting end-to-end frameworks are implemented using tools and mechanisms widely adopted within their respective research communities, enabling practical deployment across both embedded and cloud platforms.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- Boston University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Sohal, Parul
- Advisor dc:contributor.advisor
-
- Krieger, Orran
- ORCID dc:identifier.orcid
Subjects
dc:subject × 7Rights
dc:rights- Statement dc:rights
-
- Attribution 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53603