Theses and Dissertations

Item

System Support for Fine-grained Resource Management in Mobile Edge Computing

(Georgia Institute of Technology, 2024-05-16) Hsu, Ke-Jou

Multi-access edge computing (MEC) systems, similar to cloud systems, offer advantages such as multi-tenancy, fast delivery, and pay-as-you-go models. However, the limited capacity at each edge site, the collocated workloads’ stringent latency-centric performance requirements, and the heterogeneous nature of the edge, present limitations for cloud-native resource management solutions. This thesis demonstrates these limitations and addresses them via new systems support for faster and more cost-effective resource management for MEC. One limitation is the mismatch between the resource requirements for certain edge applications and the resources available at an edge site. To address this, in this thesis we first develop Couper – systems support for decomposing resource-intensive video analytics applications based on Deep Neural Networks (DNN) into finer-granular components, allowing resource management to balance the DNN inference load between the edge and the cloud, and to improve end-to-end performance. In addition, we demonstrate the importance of careful placement of components across the edge-cloud continuum. For a concrete example of a Content Delivery Network (CDN), we show that by managing the placement and collocation of components in MEC-CDN can lead to average latency reduction of 75% compared to existing solutions. We generalize the methodology used to establish this observation and develop Anitya – lightweight systems support for capturing cross-component dependencies that enables effective management of componentized microservice-based MEC application deployments. A second limitation is the mismatch among the resource allocation granularity of current MEC platforms vs. what is needed for emerging MEC workloads. We show that this gap can completely eliminate any expected edge benefits in multi-tenant settings. To address this, we develop ShapeShifter – systems support for fine-grained software-level traffic controls that augment the underlying platform capabilities to specialize the resource allocations on workload granularity. This prevents hidden congestion problems and provides 4× improvements in application performance. A third limitation is related to the mismatch among the time granularities at which cloud-native resource management operates vs. what is needed in MEC. Naive adjustments of cloud-native systems lead to prohibitive resource overheads for resource-limited MEC environments. As part of Colibri – a new observability tool for MEC, we develop new systems support for dynamically controlling and specializing the execution of control plane functionality needed for resource management, focused on resource monitoring in this case. The evaluations of the different systems developed as part of this thesis, performed using new experimental testbeds and MEC benchmarks, demonstrate that the new systems support enables improvements in the effectiveness of different resource management tasks which span the entire lifecycle of MEC application and service deployments, and results in improvements in end-to-end application performance and infrastructure efficiency.

Item

Understanding the Dynamic Operation Limits and Reliability of Highly-Scaled Silicon Germanium Heterojunction Bipolar Transistors

(Georgia Institute of Technology, ) Lee, Harrison P.

The objective of this work is to expand upon understood reliability effects in silicon- germanium (SiGe) heterojunction bipolar transistors (HBTs) and bridge the gap between a device-level understanding of reliability and system-level performance of circuits using HBTs. The aim is to better understand the safe limits of transistor biasing and investigate dynamic, transient swings past the traditional DC-defined safe operating area (SOA) of the transistor. Chapter 1 describes the need for understanding the reliability physics of SiGe HBTs. This chapter will describe how the device SOA is measured and defined and will briefly discuss the physical mechanisms at the limit of transistor operation, including avalanche multiplication, thermal runaway, and base current reversal. This chapter will also briefly discuss the design of an additional, medium breakdown (MB) transistor profile which al- lows for a designer to include aggressively scaled high speed HBTs with a higher break- down transistor in the same technology platform. Chapter 2 discusses the common damage mechanisms at play in SiGe HBTs, namely mixed-mode (MM) and high-current (HC). These two damage mechanisms occur at the outer limits of the device SOA, and understanding of these mechanisms is critical to under- stand how to maximize the potential of SiGe HBTs by operating at the edge of the device limits. Understanding damage mechanisms will also enable designers to consider the in- tended lifetime of the circuit to ensure the system meets specifications across the entire target lifetime. Chapter 3 uses pulsed-voltage measurements to investigate the dynamic breakdown of SiGe HBTs under high field. As a high voltage is applied to the collector of the device, self-heating within the transistor activates a positive feedback mechanism between device temperature, collector current, and power consumed, resulting in an uncontrollable thermal runaway. Measurements are performed by using pulsed voltage in conjunction with an emitter ballast resistor to reduce the effect of self-heating on the transistor. Results show that as self-heating is reduced, the maximum voltage limit of the device increases, and breakdown switches from being thermally dominated to impact-ionization dominated. This is significant because high-speed RF and mmWave signals switch at faster speeds than the device self-heating, and thus at high frequency, the dominant breakdown mechanism may be different from the conventional DC-defined SOA. Chapter 4 compares the performance and reliability tradeoffs of cascode amplifier cells designed using different transistor profiles. Test structures were created using the medium breakdown (MB) and high performance (HP) variants of an advanced SiGe HBT technology (GlobalFoundries 9HP). Small-signal (fT /fmax) and large-signal (gain compression) measurements are performed to quantify the performance difference between cascodes with HP and MB devices in the common base (CB) stage, and DC+RF stress measurements are used to identify the reliability difference. The results show that large cascode cells oper- ating well below the transistor peak fT show roughly equivalent performance between HP and MB, and that the cascode reliability is largely dependent upon the selected bias point and load-line swing in conjunction with the top device selection. Thus, it may be possible for circuit designers to optimize the reliability of a circuit without sacrificing performance by carefully selecting the device profile and the dynamic output swing of the cascode cell.

Item

Optimizing Of Spoiler Thickness For Total Body Irradiation (TBI) Utilizing Tomotherapy Technology

(Georgia Institute of Technology, 2024-05-13) Ziadat, Bashar

Delivering total body irradiation (TBI) with a Tomotherapy unit presents a unique challenge: minimizing lung dose while achieving a high skin dose. Acrylic spoilers offer a potential solution by slightly attenuating the incident photon beam, and therefore, reduce the lung dose. The spoilers can also increase the skin dose by producing electrons via Compton scattering interactions of the incident photon beams. This study aimed to identify the optimal spoiler thickness for clinical application for TBI with a Tomotherapy unit. The dosimetric effects of these spoilers were investigated computationally and experimentally. The results of this study show that, among the various thicknesses (0.0, 0.635, 0.953, and 1.588 cm) of the spoiler), the thickest (i.e. 1.588 cm) spoiler yielded the lowest lung dose and the highest skin dose. This conclusion is further supported by the experimentally obtained dose data from the EBT3 films that were mounted on the ArcCHECK phantom and irradiated with the Tomotherapy machine. The ion chamber measurements within the ArcCHECK system demonstrated good agreement with the TPS calculations, exhibiting a maximum error of only 0.5% for lung dose. Furthermore, the dose data measured with the ion chamber also show that lowest lung dose was achieved with the thickest spoiler.

Item

Advancing the Application of Neural Networks for Atomistic Systems in Surface Science and Catalysis

(Georgia Institute of Technology, 2024-05-10) Hu, Yuge (Nicole)

This thesis explores the integration of neural networks into studying surface science and catalysis, focusing on the development and application of neural network force fields (NNFFs). The NNFFs, once trained on quantum mechanical calculations, enable accurate prediction of energy and forces with significantly reduced computational costs, which would facilitate rapid catalyst design and optimization. A major contribution of this thesis is the development of a novel uncertainty quantification (UQ) method using the conformal prediction framework, enhancing prediction reliability across different ML models, including feed-forward and graph neural networks. Furthermore, the Python module \texttt{AmpTorch} is introduced, which integrates UQ into training NNFFs, capable of handling over 1 million configurations and diverse chemical elements. The thesis also investigates the use of NNFFs in studying platinum-graphene interactions under strain, revealing critical insights into the bonding environments, which are corroborated by Density Functional Theory calculations. Additionally, an unsupervised neural network approach is presented for selecting model compounds in catalysis, demonstrating the power of neural networks to streamline complex chemical analyses. Collectively, the thesis underscores the transformative impact of neural networks in advancing the methodological, software, and practical applications in surface science and catalysis.

Item

Topology, Geometry, and Combinatorics of Fine Curve Graphs

(Georgia Institute of Technology, 2024-05-02) Shapiro, Roberta

The goal of this thesis is to explore curve graphs, which are combinatorial tools that encode topological information about surfaces. We focus on variants of the fine curve graph of a surface, which has its vertices essential simple closed curves on the surface and whose edges connect pairs of curves that are disjoint. We will prove various geometric, topological, and combinatorial results about these curve graph variants, including hyperbolicity (or lack thereof), contractibility of induced flag complexes, automorphism groups, and admissible induced subgraphs.

Item

Back-End-of-Line Mechanical Stress Effects in SiGe HBTs at Cryogenic Temperatures

(Georgia Institute of Technology, 2024-04-30) Moody, Jackson

SiGe heterojunction bipolar transistors (HBTs) are well suited to operating in cryogenic environments, which are found in areas from deep space exploration to radio astronomy and quantum computing. One problem in deploying SiGe HBTs to cryogenic environments is increased device-to-device variability, a major factor of which is variations in mechanical stress caused by metal routing (referred to as the back-end-of-line, or BEOL) above and in the immediate vicinity of the active devices. The impact of BEOL stress on SiGe HBTs at crygenic temperatures has been presented in [1]. This work does a deeper examina- tion of the theoretical and measured impact of BEOL stress on SiGe HBTs at cryogenic temperatures.

Item

Thermophysical and Molten Salt Corrosion Behavior of Structural Materials for Next-Generation Clean Energy Systems

(Georgia Institute of Technology, 2024-05-01) Brankovic, Sonja

As next-generation clean energy technologies like concentrated solar power (CSP) and molten salt reactor (MSR) systems operate at ever higher temperatures to increase efficiency and thermal energy storage capabilities, using molten chloride salt as a heat transfer and energy storage fluid can provide many benefits, including high-temperature operation, a low operating pressure, extended storage time, and increased safety. In this extreme environment, it is essential to understand the temperature-dependent thermophysical properties and molten salt corrosion behavior of candidate structural alloys and aluminosilicate refractories for salt storage tanks, piping, and heat exchangers. Though these types of materials have been used in established applications (for example, aerospace and gas turbine engine components in the case of Ni-alloys, hot-face furnace liners for aluminosilicate refractories), corrosion studies of these types of alloys are not easily comparable in the literature; for several alloys and many of the refractories studied in this project, published corrosion data does not exist. High-temperature thermophysical data of the candidate alloys and refractories are more widely available, though not consistently in the temperature range of interest (600–800°C). The purpose of this thesis project is first to characterize the high-temperature thermophysical properties (thermal diffusivity, specific heat, and thermal conductivity) of the candidate materials. This data, combined with published results from the literature, is then used to down-select materials for temperature-dependent immersion corrosion testing. Twelve Ni- and Fe-based alloys and seven aluminosilicate refractories were initially selected for experimental testing and sourced from commercial manufacturers. The temperature-dependent thermal diffusivity and specific heat of these candidate alloys and refractories were determined via light flash analysis (LFA) and differential scanning calorimetry (DSC), respectively. These experimental results were compared with available manufacturer data of the materials’ high-temperature thermophysical properties. A subset of high-performing and commercially viable alloys and refractories, with the addition of two alumina-forming alloys, were selected for molten chloride salt corrosion testing. Samples were immersed in purified 45.98 MgCl2–38.91 KCl–15.11 NaCl wt% salt for 100 hours at 650°, 725°, and 800°C. Corrosion rates were calculated based on nominal sample densities and measured weight changes after the immersion test; comparisons of pre- and post-test surface elemental and phase compositions were performed using X-ray fluorescence (XRF) and X-ray diffraction (XRD), respectively. A more detailed cross-section analysis was performed using scanning electron microscopy (SEM) and energy dispersive X-ray spectroscopy (EDS). For the Ni-based alloys, the measured specific heat and thermal diffusivity were approximately linear as a function of temperature (which is commonly seen in manufacturer data sheets) but did see some evidence of second-order phase changes in the DSC data. In situ HT-XRD testing of several down-selected alloys showed that the alloys’ crystal structure was expanding as a function of temperature in a roughly linear manner, though there was no clear appearance of new phases or decrease in material stability. The aluminosilicate refractories exhibited no obvious phase changes in the DSC or LFA runs; this was confirmed by the in situ XRD tests. After the 100-hour immersion testing, uniform corrosion was visible on many of the samples’ surfaces and increased as a function of temperature, based on the measured mass loss of each sample. The temperature-dependent increase was most apparent in the alloys with a significant base iron content. This trend was confirmed by SEM imaging and EDS linescans of the sample cross-sections. XRF testing of the corroded alloys’ surfaces showed several compositional changes that are commonly seen in molten halide corrosion, including depletion of active metals like iron and chromium, and a corresponding enrichment in more noble elements like nickel and molybdenum. For several of the alloys, XRD testing of the corroded surfaces showed some evidence of oxide contamination. For the pre-oxidized alloys, no significant difference in performance was observed compared to the bare alloys; the developed oxide layer provided no measurable corrosion protection after 200 hours of chloride salt corrosion testing. Corrosion testing of the aluminosilicate refractories revealed no consistent, temperature-dependent trend in mass gain after 200 hours of chloride salt immersion at three temperature points. However, at higher test temperatures (725° and 800°C), vaporized chloride salt penetrated the refractories’ surface above the immersion line. A “transition line” was also observed, marking the highest level of the molten salt; this line was darker than the vapor and immersed regions of the refractories, indicating that any residue or contaminants floated on the surface of the molten salt. This thesis work is significant because it provides a broad, high-temperature thermophysical characterization of candidate alloys and aluminosilicate refractories for the next-generation solar and nuclear industries. Compared to the provided manufacturer data, the temperature-dependent runs from this thesis work provides a much finer dataset and elaborates on trends that are more subtle in the published archive. The results from the immersion molten chloride salt testing of down-selected alloys and refractories contribute important data for these same industries where understanding material corrosion resistance is critical for safe and economic performance.

Item

Design and Initial Optical Characterization of a Lean, Premixed, Prevaporized Combustor

(Georgia Institute of Technology, 2024-04-29) Budzinski, Steven Matthew

One of the challenges for wider acceptance of civil supersonic transportation (CST) is the environmental impacts of a CST fleet, especially as stricter emissions requirements are being developed. Current research utilizing a low-emissions Lean Premixed Prevaporized (LPP) concept combustor shows promising results for potential use in a larger CST engine. However, the validation of combustor stability and emissions metrics are still needed in order to properly model the combustor. This work presents the design and construction of a LPP concept combustor including components for the injector, liner, water-cooled exhaust, choke plate, and exhaust probe. After assembly, preliminary characterization of the LPP combustor is completed using digital inline holography (DIH), fuel planar laser induced fluorescence (PLIF), and OH* chemiluminescence. Due to fuel injection bias noted during the first campaign, significant effort is made to characterize the spray generated by the premixer. Here, DIH and phase Doppler particle analysis (PDPA) are compared quantitatively for the first time and used to characterize the spray as the axial injection locations, injection locations, and momentum flux ratios are varied. By presenting initial characterization results for this combustor, data is acquired to support future design improvements for the LPP concept combustor. Overall, this experiment aims to improve modeling of the LPP combustion process for future supersonic transportation applications.

Item

Transfer learning for brain signals using optimal transport

(Georgia Institute of Technology, 2024-04-29) Gupta, Ekansh

Brain-computer interfaces (BCIs) have surfaced as a powerful modality in human-machine interaction and wearable technology with powered futuristic applications like virtual reality, robot control, gaming, etc. Using BCIs, the brain's intent can be harnessed without explicit communication. Despite the vast promise, systems designed for BCIs generalize poorly to new or unseen individuals due to high variability in brain signals among different subjects, resulting in long retraining/calibration sessions. This lack of generalization is typically attributed to a covariate shift of signals in the probability space, which manifests itself as disparate marginal and class conditional distributions. In this thesis, we overview the factors contributing to poor generalization on a more granular level by analyzing a specific brain signal called the Error Potential (ErrP), a signal well-known for its noisy characteristics and high variability, explore unsupervised and semi-supervised methods to improve its generalization performance, and propose a novel algorithm to mitigate the associated covariate shift using partial target-aware optimal transport. We demonstrate our method on an ErrP dataset collected in our lab. Our method outperforms state-of-the-art models for cross-user generalization which translates to a reduction in calibration time by an order of magnitude.

Item

Do we need all the neurons we utilize? An exploration into costs and scalability for large-scale multi-relational graph learning

(Georgia Institute of Technology, 2024-04-30) Sathidevi, Lakshmi

The first step in HW-SW Co-design is algorithm selection. The algorithm chosen places hard boundaries on the kind of efficiency and scalability that can be achieved through the design of a custom hardware accelerator. A smartly selected algorithm could enable several times more performance gains than a cutting-edge hardware accelerator designed for a poorly selected algorithm. This study presents Het#GNN, an unsupervised network embedding algorithm that extends on top of the #GNN algorithm to handle multi-relational graph learning. The goal is to provide a scalable, efficient, and low-latency solution for large-scale multi-relational graph learning. Graph Neural Network (GNN) models have shown promising results in various applications. However, these models are often hindered by high memory and computational requirements, limiting their applicability to large-scale real-world scenarios. The biomedical community has shown growing interest in more efficient and less heavily parameterized methods for handling large-scale multi-relational graph learning. Het#GNN is a simple yet powerful unsupervised algorithm that addresses the efficiency and scalability challenges faced by GNN algorithms. It enables a parameter-free approach to network embedding for large-scale multi-relational graphs. Het#GNN is able to compete with the best-performing models at less than a fraction of the runtime and power costs and Het#GNN achieves this on a consumer CPU, while all other methods utilize server GPUs to accelerate their algorithm. The need for a scalable, efficient, and low-latency graph learning algorithm is increasingly important in the biomedical research community. Decagon is a very popular work that applied the multi-relational GNN model R-GCN to polypharmacy drug side-effect prediction, obtaining strong results and medically relevant predictions. However, just one epoch of training for this model takes 36 hours on a Tesla P40 GPU. This kind of turnaround time is detrimental to the pace of biomedical research and discovery. Het#GNN is presented as a solution. It is applied to the Decagon dataset to demonstrate how sometimes we maybe utilizing too many time and energy expensive neurons where none may be required. By focusing on simple yet smart choices at the algorithmic level, Het#GNN offers considerable benefits in terms of efficiency, scalability, and latency compared to other best-performing methods without requiring high-end computational resources. With this work, it is hoped that this simple and accessible contribution enables biomedical researchers to accelerate the pace of their own biomedical research and discovery pipelines that involve large-scale multi-relational graph learning. The results of this work, however, are expected to be beneficial to other domains and areas of application as well, like real-time graph learning.

Theses and Dissertations

Permanent URI for this collection

Search within this collection.

Browse

Georgia Tech Library

Browse

Recent Submissions