137091 research outputs found
Sort by
Memory System Optimizations for Parallel and Bandwidth-Intensive Workloads
No full textThe rapid proliferation of digital services—from web analytics and cloud platforms to social media and generative AI—has driven an unprecedented surge in data generation and processing demands, positioning data centers as the operational core of today’s digital economy. As hyperscale infrastructures expand to meet this exponential growth, the memory subsystem has emerged as a critical performance and cost bottleneck. Modern server nodes face mounting pressure to sustain high-capacity, high-bandwidth, and low-latency memory access as workloads increasingly rely on large in-memory datasets and parallel execution across thousands of cores. However, traditional scaling approaches—such as adding DRAM channels or relying on remote memory through RDMA—face physical, technological, and economic limitations. The resulting “memory wall” manifests as three interdependent challenges: limited capacity, constrained bandwidth, and rising latency, all intensified by the slowdown of Moore’s Law and the end of Dennard scaling.
This thesis addresses these challenges through a holistic, cross-layer co-design approach that enhances memory system performance across three hierarchical levels—chip, server, and cluster. At the chip level, HinTM introduces compiler- and hardware-assisted mechanisms to mitigate capacity aborts in Hardware Transactional Memory systems, thereby improving on-chip cache utilization and parallel execution efficiency. At the server level, SURGE dynamically harvests idle I/O bandwidth over CXL links to augment effective memory bandwidth and reduce memory access latency in bandwidth-bound workloads. Extending to the cluster scale, COMET provides a unified design-space exploration framework that jointly optimizes compute, memory, and interconnect provisioning for distributed AI and HPC workloads. Collectively, these contributions demonstrate that co-optimizing architectural mechanisms with workload and hardware characteristics can overcome the fundamental limitations of memory capacity and bandwidth scaling, enabling sustained performance improvements across modern datacenter systems
Designing ML-centric Data Systems for Efficiency and Usability
No full textOver the past six decades, relational databases have been remarkably successful in managing structured data. However, the growing demand for analytics over unstructured data, such as videos, images, and text, driven by modern Machine Learning (ML) workloads, exposes fundamental limitations in traditional database systems. Bridging this gap requires a new class of data systems that treat ML models as first-class citizens, integrating them directly into the query engine and providing optimizations tailored for their unique characteristics.
This dissertation presents the design, implementation, and evaluation of techniques that form the foundation of ML-centric data management systems. It introduces four systems—EVA, Seiden, Aero, and PRISM—that collectively address challenges of efficiency and usability across multimodal workloads.
EVA accelerates exploratory video analytics by automatically materializing and reusing the results of expensive User-Defined Functions (UDFs) through a symbolic reuse framework. Seiden revisits the long-standing “proxy model” assumption in visual databases and demonstrates that indexing directly with oracle models and exploration-exploitation sampling delivers superior execution performance and query accuracy compared to state-of-the-art video database management systems.
Aero extends Adaptive Query Processing (AQP) to ML workloads by using runtime feedback to reorder predicates and scale resources dynamically, achieving better performance improvement over static optimizers. With the advent of Large Language Models (LLMs), the interface to data systems is shifting from SQL toward natural language. While the community has largely focused on accuracy for Natural Language to SQL (NL2SQL), the dollar cost of these pipelines is often ignored. To this end, PRISM is an optimization framework that treats cost as a first-class citizen and systematically navigates the cost-accuracy trade-off tailored to each specific schema and user requirement.
Together, these contributions lay the foundation for the next generation of data systems designed for AI-driven workloads
Active, Controllable Overvoltage and Overcurrent Protection Devices for Power Systems with High Inverter-Based Resource Domination
No full textA rapid proliferation of inverter-based resources that interface distributed renewable energy resources with the bulk power system is currently underway. The changing grid paradigm with more inverter-interfaced resources poses significant challenges to modern power system protection devices. For OCP, the fault current magnitude is significantly limited and variable due to the inherent current limiting of inverters and the diversity of IBR connectivity during a fault. As a result, OCP devices must intelligently detect and trip fault currents while achieving proper protection coordination in dynamically evolving system topologies. For OVP against high energy transient surge events, such as lightning strikes, semiconductor devices in IBRs that are very sensitive to overvoltage must be protected within a small margin of the nominal ratings. Thus, a new stream of OCP and OVP solutions is needed to provide dynamic, intelligent, and active protection for IBR-dominated grid networks. This research study proposes two new concepts, the IFCM and the iMOV, which are designed to protect such systems in a decentralized fashion to avoid the need for complex, expensive, low-latency communication infrastructure. As a result, the proposed solutions provide robust, scalable, low-cost OCP and OVP solutions for the modern IBR- rich systems.Ph.D.Electrical and Computer Engineerin
Impact of Interleaved Thermally Conductive Material with Conformal Fluid Channels on Tooling Thermal Response
No full textCooling time between plastic injection and part ejection in the injection molding tooling
industry can account for 50-70% of the cycle time for molded components. It has been
determined cooling time can be decreased by 35% with conformal cooling channels and 29%
with copper inserts. The rapid development of commercially available hybrid machine tools with integrated additive and subtractive manufacturing capabilities enables the possibility of manufacturing high precision monolithic, multi-material components for the tool and die industry which utilize conformal fluid channels and copper; however, processing strategies to produce such components are not well established. This research investigates the use of
interleaved conductive material, such as copper, within a monolithic mold with integrated
conformal fluid channels for increased thermal performance of tooling. Enabled by hybrid
manufacturing, the integration of additive and subtractive processing, allows the manufacture of complex shapes with multi-material structures which are traditionally unmanufacturable with improved surface finish required by tooling applications. While the potential implementation of hybrid directed energy deposition (DED) has been acknowledged to address this need, the manufacturability of a conductive and heat treatable material interface, process planning, and tool path design considerations for varied conformal fluid channel geometries, and bi-material monolithic structure manufacturing for enhanced thermal performance manufacturability is either under-developed or not yet fully understood. This investigation began with the process-structure-property analysis of 17-4PH, pure copper, and a mixture of the two at varied ratios. The investigation continued by evaluating the interface between the two materials when 17-4PH was deposited on top of pure copper and a copper mixture. The investigation was completed when various manufacturing strategies for conformal fluid channels were conducted. With the understanding developed through each step of this investigation, an analysis of an application of a multi-material conformal fluid channel mold insert was conducted and the cool down time was nearly halved as compared to the single material insert.Ph.D.Mechanical Engineerin
Aerothermal Analysis and Design of HyperSat: An Aerobraking CubeSat With a Mechanically Deployable Heatshield
No full textConference proceedings from the AIAA SciTech Forum 2026Low-cost hypersonic platforms have been suggested as a solution to the growing need for hypersonic test data given the limitations of existing ground test techniques. As a part of these efforts, Georgia Tech and the Georgia Tech Research Institute have been developing a low-cost Hypersonic 12U CubeSat platform (HyperSat) to be used as a hypersonic testbed. HyperSat begins in a GTO orbit and uses aerobraking passes to expose the vehicle to hypersonic conditions. In order to fit within a 12U footprint and survive atmospheric heating, HyperSat utilizes a mechanically deployable heatshield with a flexible thermal protection system (FTPS). This work describes the trajectory, aerothermal analysis, and subsequent geometry modifications to ensure that the FTPS remains within heat flux material limits. The expected entry environments were bounded using a Monte Carlo trajectory simulation varying aerodynamic properties, entry conditions, atmospheric properties, and mass properties. These bounds were then utilized to perform high-fidelity aerothermal CFD simulations using LAURA and FUN3D to predict heating on the FTPS. Localized hot spots were identified at the interfaces between the deployable panels, motivating design modifications, including changes to the cone angle and nose thickness, to minimize heating.GTRI IRAD progra
Being an Open Science Champion
No full textInterview portion of Lost in the Stacks, episode 671. Features interview with Dr Lynn Kamerlin, professor of Molecular Design in the School of Chemistry and Biochemistry, and leader of the Kamerlin Lab here at Georgia Tech. She discusses how she incorporates open principles in every aspect of her research process.Interview portion of Lost in the Stacks, episode 671. Features interview with Dr Lynn Kamerlin, professor of Molecular Design in the School of Chemistry and Biochemistry, and leader of the Kamerlin Lab here at Georgia Tech. She discusses how she incorporates open principles in every aspect of her research process
Study Of Intelligent Wearables and Adaptive Systems to Measure Physiological and Physical Data for Human-Machine Interfaces
No full textThis dissertation explores the development of intelligent wearable technologies and adaptive systems that measure physiological and physical data to enhance the interaction between humans and machines. It aims to advance Human-Machine Interfaces (HMIs) through the development of wearables and systems that prioritize precision, inclusivity, and user adaptability. The first study focuses on an Electrooculography (EOG) driven headband, a wearable HMI designed for precise and continuous control through non-invasive eye biopotential tracking. This system integrates advanced signal processing and real-time classification algorithms through mobile to address challenges such as noisy biopotential signals and inconsistent wireless control., demonstrating its utility in telesurgery and virtual reality applications. The second study introduces a Two-Camera Eye-Tracking System (TCES), a novel dual-camera setup optimized for eye movement and gaze tracking. By addressing anatomical variability and environmental challenges, the system ensures high accuracy and inclusivity, enabling precise control of any screen-based tasks and external controllers, including robotic arms. The application of advanced machine learning algorithms and data-driven control further enhances its adaptability, making it accessible for people, including individuals with physical disabilities. Finally, the third study presents a Multiplexed Sensing Suit, a full-body motion tracking system that integrates multi-sensor arrays with cloud-based machine learning system. This platform enables real-time monitoring and classification of complex human motions, with applications ranging from sports analytics to rehabilitation. By combining flexibility, scalability, and detailed motion analysis, this study advances the field of wearable HMI beyond its capabilities. Through these contributions, this dissertation aims to bridge the gap between user intent and machine response, creating systems that are both practical and transformative. By integrating advanced sensing systems, adaptive algorithms, and scalable real-time infrastructure, the research highlights the potential for HMI’s promotion to foster continuous, natural, and user-centric interactions in everyday and professional settings.Ph.D.Electrical and Computer Engineerin
Mechanically Intelligent Elongate Limbless Robots for Locomotion in Complex Land and Water Environments
No full textLimbless animals such as snakes and nematodes exhibit remarkable capability in navigating complex environments, inspiring the development of limbless robotic systems. However, most existing designs consist of rigid segments actuated by rotational motors and often face limitations in mobility and adaptability within heterogeneous or unstructured terrains. This thesis introduces a new design paradigm centered on mechanical intelligence (MI). A novel actuation mechanism is presented, featuring bilateral actuation along a flexible spine that models the musculoskeletal systems of animals. This mechanism enables effective open-loop locomotion in complex environments through the exploitation of passive body mechanics and body-environment interactions, thereby reducing reliance on complex control algorithms while guaranteeing adaptability. Building on this foundation, computational intelligence (CI) techniques such as gait optimization, tactile sensing, and closed-loop control are incorporated to achieve enhanced performance across both terrestrial and aquatic environments. The thesis is organized around three aims: (1) to develop bilaterally actuated limbless robots to identify and quantify the principles of MI, (2) to design and optimize gaits that take advantage of MI for improved performance, and (3) to extend established MI principles and discover new ones in aquatic environments, exploring the synergy between MI and CI for robust, adaptive amphibious autonomy. This work contributes to the development of versatile limbless robots with enhanced autonomy and resilience, supporting applications in search-and-rescue operations, industrial inspections, precision agriculture, and planetary exploration
Optimization of Lens-Based Antenna Arrays Leveraging Optical-mmWave Convergence (ROF) for Simplified Base Stations Enabling Ultra Dense Wireless Networks
No full textThe ever-increasing demands for higher wireless capacity, enhanced coverage, and support for massive network traffic have driven research towards ultra-dense wireless networks (UDNs). However, realizing truly ultra-dense deployment of base stations is challenging for a variety of reasons. Some of the main inhibitors are the high complexity, power consumption and cost of conventional base stations, which rely on a large number of high-power, bulky, expensive and narrow-band components. These hardware requirements increase both capital expenditure (CAPEX) and operational expenditure (OPEX), making large-scale deployment difficult. Beyond this, the continual evolution of wireless standards and addition of spectrum demands flexible architectures, which can be upgraded and reconfigured easily without significant build-out costs.
This thesis presents simplified base-station architectures enabled by optical and millimeter-wave (mmWave) convergence leveraging passive, compact beamforming networks and wideband or multi-band antenna arrays. By centralizing processing and computationally heavy processes at the central office, or hub location, broadband mmWave signals are transmitted via optical fiber to simplified remote radio units (RU). At the RU, only passive, low complexity beamforming hardware, optical-electrical converters (OE) and amplifiers are needed in the downlink, significantly reducing hardware count, size and power consumption demand.
The optical fronthaul is modeled in VPIphotonics TransmissionMaker and simulations were performed. Complementary electromagnetic simulations of the Rotman lens and antenna arrays in CST Microwave Studio verify multi-beam and multi-band/wideband performance. The combined optical-RF simulation framework demonstrates consistent beam steering over multi-band/wideband frequency ranges and effective, transmission of multi-GHz-bandwidth signals over fiber links, within standard limits.
The results confirm that the converged optical/mmWave architecture offers a scalable, energy-efficient, and cost-effective pathway for next-generation networks. By simplifying the radio unit and centralizing complexity, the proposed system lowers CAPEX, OPEX, and network build-out costs while maintaining high throughput and flexibility. This work provides a foundation for future 6G-ready, software-defined, remote-controlled, and reconfigurable radio access networks, addressing the core challenges of capacity, cost, and deployment scalability in ultra-dense wireless environments
Autonomous System for Identifying and Capturing Floating Waste
No full textPlastic pollution in aquatic environments poses a critical challenge, causing significant harm to ecosystems and human health. Traditional cleanup methods are often labor-intensive and limited in scalability and coverage, prompting interest in autonomous robotic solutions. This dissertation investigates the development of an autonomous system for detecting and collecting floating waste, combining advanced visual perception and control based on reinforcement learning (RL) into an integrated system tailored for aquatic environments. The application not only addresses an important ecological problem but also presents substantial technical challenges in perception, navigation, and field deployment.
The first contribution of this work is the enhancement of visual perception through polarimetric imaging. Specular reflections from water surfaces complicate detection tasks using standard cameras. To overcome this, the PoTATO dataset was created, containing images of floating plastic bottles captured with a polarimetric camera. The dataset includes multiple polarimetric modalities, and experiments demonstrate that polarimetric inputs, especially when fused with color information, improve detection accuracy under diverse lighting conditions.
The second contribution is the development of a reinforcement learning framework for autonomous surface vehicle (ASV) navigation, incorporating hydrodynamic modeling and domain randomization to reduce the sim-to-real (simulation-to-reality) gap. The RL-based controller is evaluated with respect to energy efficiency and task completion time, and is tested under disturbances such as shifting payloads and asymmetric drag. Comprehensive field experiments demonstrate the robustness of the trained policies and validate the training approach.
Finally, the dissertation integrates perception and control into an end-to-end pipeline for autonomous waste collection. Detected objects are mapped to navigation targets using calibrated transformations, and the ASV is guided by the RL agent. The integrated system is evaluated in both simulation and real-world settings, highlighting the feasibility for environmental cleanup applications.
In summary, this work demonstrates a comprehensive approach to autonomous waste collection by advancing perception, control, and integration methods. Field trials are central to the evaluation, ensuring that the system performs reliably under real-world conditions. These findings highlight the potential of autonomous systems to contribute to environmental monitoring and waste mitigation