UWSpace

UWSpace is the University of Waterloo’s institutional repository for the free, secure, and long-term home of research produced by faculty, students, and staff.

Depositing Theses/Dissertations or Research to UWSpace

Are you a Graduate Student depositing your thesis to UWSpace? See our Thesis Deposit Help and UWSpace Thesis FAQ pages to learn more.

Are you a Faculty or Staff member depositing research to UWSpace? See our Waterloo Research Deposit Help and Self-Archiving pages to learn more.

Photo by Waterloo staff

Communities in UWSpace

Select a community to browse its collections.

Recent Submissions

  • Thumbnail Image
    Item type: Item ,
    Efficiently Training Deep Learning Models on Elastic and Heterogeneous Cloud Resources
    (University of Waterloo, 2026-05-12) Guo, Runsheng
    Deep Neural Networks (DNNs) have demonstrated remarkable success across diverse domains, but their training requires substantial computational resources and is typically parallelized across large GPU clusters. However, such clusters are prohibitively expensive for most organizations to own and manage. Hence, instead of owning and managing their own clusters, organizations often rent clusters on cloud platforms to meet their training needs. While cloud environments offer elastic scalability and heterogeneous hardware options, they also introduce significant challenges for efficient distributed DNN training. Specifically, existing training frameworks lack support for dynamic reconfiguration during training, limiting the exploitation of cloud elasticity. Additionally, most systems assume homogeneous clusters, which rarely reflect the heterogeneous GPU clusters that organizations commonly use due to hardware availability constraints. Furthermore, heterogeneous network conditions in cloud environments create communication bottlenecks that limit the scalability of existing approaches. This thesis presents three systems that collectively address these limitations to enable efficient distributed DNN training on elastic and heterogeneous cloud resources. First, Hydrozoa leverages cloud elasticity through serverless containers, enabling dynamic scaling and configuration changes during training without the traditional pitfalls of serverless computing. By combining data and model parallelism with fine-grained resource provisioning, Hydrozoa achieves cost-effective training while eliminating cluster management overhead. Second, Cephalo addresses heterogeneous GPU clusters by independently balancing compute and memory resources across GPUs with different capabilities. Unlike existing approaches that tie workload assignment to computational speed, Cephalo separately optimizes compute distribution through proportional batch sizing and memory utilization through intelligent partitioning of training state, activation checkpointing, and gradient accumulation strategies. Third, Zorse tackles heterogeneous network conditions, which are particularly common in heterogeneous clusters, by efficiently combining memory-efficient data parallelism with pipeline parallelism. Through interleaved pipelining, parameter and activation offloading, and heterogeneous pipeline parallelism configurations, Zorse achieves both communication and memory efficiency for training large DNN models across diverse network topologies. The experimental evaluation demonstrates that these systems significantly improve training efficiency and resource utilization compared to existing approaches. Hydrozoa reduces training costs while providing seamless scalability, Cephalo simultaneously achieves high compute and memory utilization in heterogeneous clusters, and Zorse maintains high throughput under varying network conditions. Together, these contributions make distributed DNN training more accessible, cost-effective, and efficient in modern cloud environments, advancing the state of the art in large-scale machine learning infrastructure.
  • Thumbnail Image
    Item type: Item ,
    Evaluating Mercury Deposition over Space and Time in Critical Nesting Habitat for Endangered Whooping Crane (Wood Buffalo National Park)
    (University of Waterloo, 2026-05-12) Lacey, Amy Lynn
    The only remaining wild, self-sustaining population of endangered Whooping Crane (the Aransas-Wood Buffalo population of Grus americana) nests and breeds in a remote boreal landscape of northwestern Canada known as the Whooping Crane Summer Range (WCSR), a Ramsar Wetland of International Importance. Although geographically isolated from industrial development, ponds in the WCSR may receive deposition of substances of concern, including mercury (Hg), from far-field sources via direct atmospheric transport and deposition and locally via remobilization of legacy Hg stored in soils and catchments. Mercury is of particular concern because it is a potent neurotoxin capable of long-range transport and accumulation in aquatic ecosystems. However, there are no data available to evaluate present-day Hg concentrations in aquatic sediment, the enrichment of Hg relative to naturally occurring baseline concentrations, or long-term accumulation of Hg within WCSR pond sediment. This study investigates total mercury (THg) concentrations over space and time in pond sediments of the WCSR to determine whether the concentrations pose risk of adverse biological effects on whooping crane and other aquatic biota and to quantify the enrichment and storage of THg in pond sediment. Total Hg concentrations were measured in surface sediments collected in 2024 from 63 ponds and in radiometrically dated sediment cores spanning the past ~370 years from three ponds within the Sass–Klewi Nesting Area (SKNA), a subregion of the WCSR. Sediment cores were used to establish a multi-site pre-industrial baseline to quantify enrichment using organic matter–normalized Enrichment Factors (EFs), and to calculate cumulative inventories of excess THg for comparison with other Canadian lakes at near- and far-field distances from major sources of emissions. Concentrations of THg ranged from 2.7 – 45.4 ng g⁻¹ (mean 20.0 ± 9.9 ng g⁻¹, n = 63) in the surface sediments and from 5.4 – 72.0 ng g⁻¹ in the sediment cores. All concentrations are well below the Canadian Interim Sediment Quality Guideline of 170 ng g⁻¹ and indicate that adverse biological effects on aquatic life are unlikely to occur at present and in the past. Strong, positive linear relations between THg concentrations and organic matter (OM) content in sediment deposited before ~1900 at three ponds allowed for construction of a multi-site pre-industrial OM-normalized baseline capable of estimating the amount of enrichment in samples deposited since 1900 in sediment cores each of the three pond and the surface sediment samples from 63 ponds. Peak THg enrichment occurred between ~1960 and ~2010 at the three ponds, when EFs range from 1.5-2.6 and signify ‘minimal enrichment’ based on widely used categorization. After ~2010, EFs decline rapidly to around 1.0, indicating a return to concentrations that existed before 1900. The mean EF for the surface sediment samples collected from 63 ponds in 2024 is 1.03, indicating no enrichment relative to pre-1900 sediment, and EFs exceed the 1.5 threshold for ‘minimal enrichment’ at fewer than 10% of the ponds. The cumulative inventory of excess THg at pond SK 43 falls within a narrow range reported for other lakes in Canada at far-field distances from industrialization and is 700 times lower than a lake at near-field distance from a major point source of Hg emissions (mine and smelter at Flin Flon, Manitoba). These findings suggest that mercury poses little risk of harm to whooping crane and other aquatic organisms. The pre-1900 baselines generated during the research can be used in future sediment quality monitoring to detect whether ongoing climatic and environmental changes begin to elevate THg concentrations above natural levels.
  • Thumbnail Image
    Item type: Item ,
    Design, Testing, and Analysis of Advanced Massive MIMO Transmitters
    (University of Waterloo, 2026-05-12) Lim, Jin Gyu
    The evolution of Fifth Generation (5G) wireless systems is driven by the need for higher data rates, lower latency, and improved coverage. Massive Multiple Input, Multiple Output (MIMO) transmitter front-ends have emerged as a key enabling technology, employing multiple parallel transmitter chains—each comprising Digital Signal Processing (DSP), Power Amplifiers (PAs), and antenna elements—to exploit spatial multiplexing and enable multi-user beamforming, at the cost of increased system complexity. The PA is a critical component in each transmitter chain, as its efficiency dominates energy consumption and its output power determines coverage. However, PA performance is highly sensitive to load impedance. While isolators in 4G systems maintain a constant 50 Ω load, their use in massive MIMO arrays is impractical due to cost, size, bandwidth, and integration constraints. Consequently, antenna mismatches and mutual coupling introduce dynamic load variations that degrade PA and overall system performance, motivating a holistic system-level design approach. The first objective of this thesis is to develop tools for system-level analysis of massive MIMO transmitters under realistic excitation. A multidisciplinary co-simulation frame- work integrating DSP, Radio Frequency (RF), and electromagnetic domains is proposed to capture signal processing, PA nonlinearities, and antenna coupling within a unified environment. Experimental validation using a four-channel fully digital MIMO transmit- ter demonstrates accurate prediction of system-level trends. A sixteen-channel testbed is further developed to validate design strategies and capture hardware-specific effects. The second objective is to enable PA design under realistic system conditions. To mitigate the computational complexity of large-scale simulations, an emulation platform is developed that reproduces massive MIMO loading conditions using a single PA. This approach enables efficient characterization and optimization under dynamic impedance environments. Combined with the co-simulation framework, it supports PA design directly at the system level rather than under idealized 50 Ω assumptions. The third objective is to investigate the impact of precoding on PA behavior. Con- ventional Digital Pre-Distortion (DPD) linearizes PAs using uncorrelated signals prior to precoding, implicitly assuming invariant load conditions. However, precoding alters signal correlation and power distribution, thereby modifying the load impedance seen by each PA in the presence of mutual coupling, which degrades linearization performance. To ad- dress this, an alternative architecture is explored in which precoding precedes linearization, enabling improved robustness and reduced DPD complexity under dynamic conditions.
  • Thumbnail Image
    Item type: Item ,
    Monolithic Integration of GeTe Switches and BST Varactors for Reconfigurable Millimeter-Wave Devices
    (University of Waterloo, 2026-05-12) Golcheshmeh, Mehran
    Reconfigurable microwave and millimeter-wave systems require tunable components that provide low loss, high resolution, and compact implementation. Conventional approaches based on semiconductor devices, microelectromechanical systems (MEMS), and purely analog or digital tuning techniques face limitations in loss, tuning range, resolution, and integration complexity. Therefore, alternative approaches are needed that can overcome these challenges while remaining compatible with integrated fabrication processes. This thesis presents the development of tunable RF components based on the monolithic integration of ferroelectric barium strontium titanate (BST) varactors and phase-change material (PCM) germanium telluride (GeTe) switches. BST varactors provide continuous analog tuning with low power consumption, while GeTe switches enable discrete, nonvolatile reconfiguration. The combination of these technologies enables a hybrid analog–digital tuning approach that improves tuning range and flexibility. The work begins with the development and optimization of fabrication process for BST thin-film varactors, followed by their application in tunable circuits. A monolithic fabrication process is then developed to integrate BST varactors and GeTe switches. The challenges associated with material compatibility and process conditions are addressed, and both BST and GeTe devices are fabricated and characterized. Using this platform, hybrid analog–digital varactors with enhanced tuning range are demonstrated. Finally, the hybrid tuning approach is applied to the design and implementation of phase shifters, including true-time-delay (TTD) and reflective-type architectures. These designs combine the advantages of analog and digital tuning to achieve improved phase control, compact implementation, and reduced loss compared to conventional approaches. The results presented in this thesis demonstrate the effectiveness of combining BST varactors and GeTe switches for the realization of reconfigurable millimeter-wave components, providing a practical approach for next-generation tunable RF systems.
  • Thumbnail Image
    Item type: Item ,
    Decentralized and Agentic Spectrum Management in Cognitive Wireless Networks
    (University of Waterloo, 2026-05-11) Abognah, Anas
    Dynamic spectrum management and sharing have been the subject of extensive research and development for many years. The ever-increasing demand for wireless spectrum from an exponentially growing number of devices and applications has led to a spectrum scarcity problem that remains unsolved. In addition, the rigid and prolonged nature of the regulatory processes of manually allocating spectrum has led to large swaths of spectrum bands being underutilized and inaccessible to new applications. Dynamic spectrum sharing can alleviate these problems by enabling new applications and devices to opportunistically access unused spectrum. Multiple spectrum sharing frameworks have been proposed by regulatory bodies where access to the shared spectrum is controlled and managed by a centralized third-party controller. However, these centralized spectrum sharing frameworks fail to provide truly dynamic and scalable spectrum sharing as they lack mechanisms for spectrum trading and do not provide incentives for primary users to participate in such models. In addition, existing decentralized spectrum management approaches rely on numerical optimization models that lack autonomous decision making capabilities, and are semantically blind and unable to interpret the unstructured regulatory policies and requirements. The need for a fully dynamic, and autonomous, spectrum sharing framework that satisfies the regulatory requirements and provides built-in economic incentives still exists. In this thesis, we propose and implement a fully decentralized spectrum management and sharing framework that resolves the issues inherent in the centralized model and closes the semantic gap through autonomous cognitive agents. We implement a comprehensive decentralized model that converges blockchain technology, federated learning, and Large Language Model (LLM) agents to automate and optimize dynamic spectrum sharing, sensing, and access in a single framework. The implemented model eliminates the reliance on centralized brokers through a two-tier Hyperledger Fabric blockchain network that guarantees trust, transparency, and immutable audit trails for spectrum sharing while eliminating single points of failure. In addition, the model facilitates cooperative decentralized spectrum sensing via federated model training on the blockchain achieving 92% detection accuracy. Finally, we implement BLAST (Blockchain LLM Agentic Spectrum Trading), which eliminates static decision-making and requirements analysis through autonomous cognitive agents. We demonstrate that LLM-driven agents employing game-theoretic reasoning within second-price sealed-bid auctions maximize social welfare and spectrum allocation efficiency and significantly outperform traditional heuristic strategies and state-of-the-art non-LLM decentralized models. This research establishes a concrete architectural blueprint for 6G and beyond, where decentralized intelligence, economic incentives, and regulatory compliance coexist within a unified, autonomous execution framework.