Some implementation characteristics of the Independent Modal-Space Control method are considered. It is shown that the control method is completely robust with respect to modeling errors and plant truncation effects. The globally optimal control of distributed systems requires distributed actuators. Instead of using distributed actuators, the distributed control is approximated with discrete actuators. Since the distributed control is closely approximated, the closed-loop poles are computed as a perturbation of the distributed control. The discrete actuators are located such that the control spillover is minimized. / M.S.

As distributed generations and renewable energy are becoming the fastest growing segment of the energy industry, the technical issues and environmental impacts have to be studied and understood. The large number of small-scale Microgrid components with their own characteristics is a big challenge for Microgrid modeling, simulation, planning and operation. The major goal of this thesis is to build a library of various Microgrid components. First of all, the thesis is going to present a detailed description of Microgrid models with moderate complexity. Next, it will present the modeling of loads, utility grid and transmission lines. Then, the paper will discuss the distributed generation models that have been developed in Matlab/Simulink including Diesel Engine, Fuel Cell, Micro Gas Turbine, Wind Turbine, Photovoltaic Cell, along with the detailed modeling of short-term storage (Battery, Pumped Hydro Storage, Flywheel, and Supercapacitor). In addition to steady-state study, the thesis will also discuss the hybrid sample systems that are built to investigate their transient responses. To enhance the simulation performance, some improvements on modeling and simulation will be introduced as well. To accommodate the high demand of renewable energy and the environment policy, the planning and operation the of Micro-source generators has been studied using HOMER. Simulation results show a case study of an optimal microgrid configuration on Ontario area in Canada. Sensitivity variables are specified to examine the effect of uncertainties, especially in a long-term planning. Also, demand side management plays an important role in the operation of Microgrid. Based on raw data, case studies are carried out to investigate and validate the demand response methods. Finally, the philosophy for Microgrid protection, especially Time-delay overcurrent protection, will be briefly introduced in both gird-connected and islanding modes. / Master of Science

Distributed deep learning has achieved tremendous success for solving scientific problems in research and discovery over the past years. Deep learning training is quite challenging because it requires training on large-scale massive dataset, especially with graphics processing units (GPUs) in latest high-performance computing (HPC) supercomputing systems. HPC architectures bring different performance trends in training throughput compared to the existing studies. Multiple GPUs and high-speed interconnect are used for distributed deep learning on HPC systems. Extant distributed deep learning systems are designed for non-HPC systems without considering efficiency, leading to under-utilization of expensive HPC hardware. In addition, increasing resource heterogeneity has a negative effect on resource efficiency in distributed deep learning methods including federated learning. Thus, it is important to focus on an increasing demand for both high performance and high resource efficiency for distributed deep learning systems, including latest HPC systems and federated learning systems. In this dissertation, we explore and design novel methods and frameworks to improve resource efficiency of distributed deep learning training. We address the following five important topics: performance analysis on deep learning for supercomputers, GPU-aware deep learning job scheduling, topology-aware virtual GPU training, heterogeneity-aware adaptive scheduling, and token-based incentive algorithm. In the first chapter (Chapter 3), we explore and focus on analyzing performance trend of distributed deep learning on latest HPC systems such as Summitdev supercomputer at Oak Ridge National Laboratory. We provide insights by conducting a comprehensive performance study on how deep learning workloads have effects on the performance of HPC systems with large-scale parallel processing capabilities. In the second part (Chapter 4), we design and develop a novel deep learning job scheduler MARBLE, which considers efficiency of GPU resource based on non-linear scalability of GPUs in a single node and improves GPU utilization by sharing GPUs with multiple deep learning training workloads. The third part of this dissertation (Chapter 5) proposes topology-aware virtual GPU training systems TOPAZ, specifically designed for distributed deep learning on recent HPC systems. In the fourth chapter (Chapter 6), we conduct exploration on an innovative holistic federated learning scheduling that employs a heterogeneity-aware adaptive selection method for improving resource efficiency and accuracy performance, coupled with resource usage profiling and accuracy monitoring to achieve multiple goals. In the fifth part of this dissertation (Chapter 7), we are focused on how to provide incentives to participants according to contribution for reaching high performance of final federated model, while tokens are used as a means of paying for the services of providing participants and the training infrastructure. / Doctor of Philosophy / Distributed deep learning is widely used for solving critical scientific problems with massive datasets. However, to accelerate the scientific discovery, resource efficiency is also important for the deployment on real-world systems, such as high-performance computing (HPC) systems. Deployment of existing deep learning applications on these distributed systems may lead to underutilization of HPC hardware resources. In addition, extreme resource heterogeneity has negative effects on distributed deep learning training. However, much of the prior work has not focused on specific challenges in distributed deep learning including HPC systems and heterogeneous federated systems, in terms of optimizing resource utilization.This dissertation addresses the challenges in improving resource efficiency of distributed deep learning applications, through performance analysis on deep learning for supercomputers, GPU-aware deep learning job scheduling, topology-aware virtual GPU training, and heterogeneity-aware adaptive federated learning scheduling and incentive algorithms.

Coulomb propulsion has been proposed for spacecraft cluster applications with separation distances on the order of dozens of meters. This thesis presents an investigation of analytic charge solutions for a planar and three dimensional four satellite formations. The solutions are formulated in terms of the formation geometry. In contrast to the two and three spacecraft Coulomb formations, a four spacecraft formation has additional constraints that need to be satisfied for the individual charges on the spacecraft to be unique and real. A spacecraft must not only satisfy the previously developed inequality constraints to yield a real charge solution, but it must also satisfy three additional equality constraints to ensure the spacecraft charge is unique. Further, a method is presented to reduce the number of equality constraints arising due the dynamics of a four spacecraft formation. Formation geometries are explored to determine the feasibility of orienting a square formation arbitrarily in any given plane. The unique and real spacecraft charges are determined as functions of the orientation of the square formation in a given principal orbit plane. For a three-dimensional tetrahedron formation, the charge products obtained are a unique set of solution. The full three-dimensional rotation of a tetrahedron is reduced to a two angle rotation for simpler analysis. The number of equality constraints for unique spacecraft charges can not be reduced for a three-dimensional formation. The two angle rotation results are presented for different values of the third angle. The thesis also presents the set up for a co-linear four-craft problem. The solution for the co-linear formation is not developed. The discussion of co-linear formations serves as an open question on how to determine analytic solutions for system with null-space dimension greater than 1. The thesis also presents a numerical tool for determining potential shapes of a static Coulomb formation as a support to the analytical solutions. The numerical strategy presented here uses a distributed Genetic Algorithm (GA) as an optimization tool. The GA offers several advantages over traditional gradient based optimization methods. Distributing the work of the GA over several processors reduces the computation time to arrive at a solution. The thesis discusses the implementation of a distributed GA used in the analysis of a static Coulomb formation. The thesis also addresses the challenges of implementation of a distributed GA on a computing cluster and presents candidate solutions. / Master of Science

In the last decade the world has seen a steadily increasing number of Cube Satellites deployed to Low Earth Orbit. Traditionally, these cubesats rely on Amateur Radio communications technology that are proven to work from space. However, as data volumes increase, the existing Amateur Radio protocols, combined with the restrictions of use for the Amateur Radio Spectrum, as well as the trend to build one control station per cubesat, result in a bottle neck effect whereby existing communications methods are no longer sufficient to support the increasing data volumes of the spacecraft. This Masters Thesis work explores the concept of deploying a network of distributed ground station receiver nodes for the purposes of increasing access time to the spacecraft, and thereby increasing the potential amount of data that can be transferred from orbit to the ground. The current trends in cubesat communications will be analyzed and an argument will be made in favor of transitioning to more modern digital communications approaches for on orbit missions. Finally, a candidate ground station receiver node design is presented a possible design that could be used to deploy such a network. / Master of Science

Mobile Ad hoc NETworks (MANETs) are networks in which the participating nodes can move freely without having to worry about maintaining a direct connection to any particular fixed access point. In a MANET, nodes collaborate with each other to form the network and as long as a node is in contact with any other member of the network, it—at least in theory—is part of the network and can communicate with all other nodes. An important function of network management is to observe current network conditions: at the node level, this may mean keeping track of arriving and departing traffic load; at the network level, the system must monitor active routes and changes in network topology. In this research, we present the design and implementation of a distributed network monitoring system for MANETs. Our system is completely distributed, generates no additional traffic on the network and produces a dynamic picture of the network level and node level information on a graphical user interface. In our proposed scheme, multiple monitoring nodes collaborate to achieve a reasonably accurate snapshot of the network conditions. These monitoring nodes passively sniff network traffics and gather information from the network to construct partial network views. They then transmit their findings to a management unit where these local views are put together to produce a comprehensive picture of the network. The communication between all management nodes (a monitoring unit and a management node) takes place in an out-of-band communication link. Therefore, our monitoring solution does not depend on the MANET to perform, hence is robust to network partitioning, link breaks, node's death and node misbehavior in the monitored MANET. Our solution provides a snapshot of the network topology that includes information about node-level behavior ratings and traffic activity. The information provided by our monitoring system can be used for network management as well as for security assessment, including anomaly detection. Information regarding individual nodes' behavior can be used for detecting selfishness in the network. Also, an approximation of arriving and departing traffic levels at each node is important in the context of quality of service, load balancing and congestion control. Furthermore, the network topology picture can provide valuable information to network management in detecting preferred routes, discovering network partitioning and in fault detection. We developed a proof-of-concept implementation of our system, which works with the Optimized Link State Routing (OLSR) protocol. Through experimental studies with up to 10-node MANETs, we were able to determine the feasibility and workability of our system. The scheme proved to be robust with respect to mobility, rapid changes in the network topology and node connectivity. Throughout our experiments we observed that our system replicated changes in the network on the GUI with less than two seconds delay. Also, when deployed in a high-traffic environment, with multiple TCP and UDP flows throughout the network, the system was able to report traffic load on each node accurately and consistently. On average, CPU consumption on monitoring nodes was about 3.5% and the GUI never took up more than 4% of the processing power (general-purpose laptop computers were used throughout the experiments). Also, the overall storage capacity needed for archiving the information files was estimated as 1 Mbytes for monitoring a 10-node MANETs for 30 minutes. Unobtrusive and distributed nature of our proposed approach helps the system to adapt to the constantly changing nature of MANETs and be able to provide valuable network management, security assessment and traffic analysis services, while requiring only modest processing and storage resources. The system is capable of quickly responding to changes in the network and is non-intrusive, generating no additional traffic on the MANET it monitors. / Master of Science

Intrusion detection systems are an integral part of web application security. As Internet use continues to increase, the demand for fast, accurate intrusion detection systems has grown. Various IDSs like Snort, Zeek, Solarwinds SEM, and Sleuth9, detect malicious intent based on existing patterns of attack. While these systems are widely deployed, there are limitations with their approach, and anomaly-based IDSs that classify baseline behavior and trigger on deviations were developed to address their shortcomings. Existing anomaly-based IDSs have limitations that are typical of any machine learning system, including high false-positive rates, a lack of clear infrastructure for deployment, the requirement for data to be centralized, and an inability to add modules tailored to specific organizational threats. To address these shortcomings, our work proposes a system that is distributed in nature, can actively learn and uses experts to improve accuracy. Our results indicate that the integrated system can operate independently as a holistic system while maintaining an accuracy of 99.03%, a false positive rate of 0.5%, and speed of processing 160,000 packets per second for an average system. / Master of Science / Intrusion detection systems are an integral part of web application security. The task of an intrusion detection system is to identify attacks on web applications. As Internet use continues to increase, the demand for fast, accurate intrusion detection systems has grown. Various IDSs like Snort, Zeek, Solarwinds SEM, and Sleuth9, detect malicious intent based on existing attack patterns. While these systems are widely deployed, there are limitations with their approach, and anomaly-based IDSs that learn a system's baseline behavior and trigger on deviations were developed to address their shortcomings. Existing anomaly-based IDSs have limitations that are typical of any machine learning system, including high false-positive rates, a lack of clear infrastructure for deployment, the requirement for data to be centralized, and an inability to add modules tailored to specific organizational threats. To address these shortcomings, our work proposes a system that is distributed in nature, can actively learn and uses experts to improve accuracy. Our results indicate that the integrated system can operate independently as a holistic system while maintaining an accuracy of 99.03%, a false positive rate of 0.5%, and speed of processing 160,000 packets per second for an average system.

The integration of distributed generators (DGs) into the distribution grid has exacerbated voltage unbalance issues leading to greater risks of reducing equipment lifetime, equipment damages, and increased ohmic losses. Most approaches to regulating voltage in distribution systems only focus on voltage magnitude and neglect phasor discrepancies and do little to remedy voltage unbalance. To combat this, a novel Optimal Power Flow (OPF) is designed to help operate these resources in a manner that curtails voltage unbalance using the reactive power compensation capabilities of inverters. The OPF was run for a wide variety of loading conditions on a pair of systems using MATLAB and was shown to improve the voltage profile of the system in addition to minimizing losses in most cases. However, it is noted that the OPF loses exactness in highly stressed conditions and is unable to provide meaningful solutions / Master of Science / With the power grid getting greener and smarter by the day, a slew of new challenges arise to overcome. Distributed sources of energy like solar panels and batteries are being added to the grid right from the household level. While they are desirable for reducing our need for traditional sources of energy, the addition of these resources has been shown to cause issues in the quality of the power grid. This is particularly observed at the low-voltage domestic part of the grid where the resources cause issues with the voltage quality. The distribution grid is unbalanced by nature and adding these resources only amplifies this problem. To help mitigate voltage quality issues grid operators are starting to require voltage regulation capabilities from resources to be connected to the grid and a lot of work has been conducted to find the optimal strategies for operating these resources. However, existing paradigms for these sources only focus on fixing the voltage magnitude part of the power quality and neglect phasor relationships. This thesis aims to bridge this gap by developing a method to determine the optimal operation of these resources by using the voltage regulation capability to address both voltage magnitude and voltage unbalance issues in addition to optimal operation.

Acoustic Frequency Domain Reflectometry (AFDR) is a novel technique employing frequency modulated continuous wave (FMCW) methods in solid acoustic waveguide reflectometry. It is particularly suited to dispersion compensation and phase compensation due to the measurement domain being the frequency domain. This work rigorously analyzes, develops, and experimentally demonstrates AFDR, alongside various compensation methods and demodulation techniques. Distributed measurement of temperature is tested using several novel signal processing algorithms for strain determination and is estimated to have a resolution of 0.58 °C over a 20 cm gauge length. An error correction algorithm to improve SNR in the measurement of strain is proposed and validated. The sensing system has a theoretical spatial resolution of 2 mm and an estimated sensing resolution limit of about 1 cm. AFDR and the associated signal processing developments are positioned to be transformative across many areas of acoustics, with significant potential for distributed sensing along an acoustic waveguide. / Doctor of Philosophy / Acoustic Frequency Domain Reflectometry (AFDR) is demonstrated as a novel method for using acoustic waves to sense different material parameters. Acoustic waves can be guided down various structures, such as a metal wire. Rather than sending out a short burst of acoustic power and analyzing its echoes in the metal wire, this technique uses a constant source of acoustic waves with varying frequency, instead recording how the electrical characteristics of the acoustic source change as frequency changes. Since the measurement is made across frequency, this method is particularly suited to correct for various aspects of the acoustic wave that change with frequency in an otherwise undesirable way. The ability to compensate for acoustic wave speeds that change with frequency as well as imperfections intrinsic to the tuning itself using multiple new methods is demonstrated. Distributed measurement of temperature is tested using various signal processing algorithms, and estimated to have a resolution of 0.58 °C for a 20 cm sensing length. The validated sensing system theoretically has the ability to resolve changes over 2 mm, and the resolution over which sensing may be possible is estimated to be 1 cm. AFDR and the associated signal processing developments are positioned to be transformative across many areas of acoustics, with significant potential for distributed sensing along an acoustic waveguide.

Wireless distributed computing networks (WDCNs) will become the next frontier of the wireless industry as the performance of wireless platforms is being increased every year and wireless industries are looking for "killer" applications for increased channel capacity. However, WDCNs have several unique problems compared with currently well-investigated methods for wireless sensor networks and wired distributed computing. For example, it is difficult for WDCNs to be power/energy efficient considering the uncertainty and heterogeneity of the wireless environment. In addition, the service model has to take account of the interference-limited feature of wireless channels to reduce the service delay. Our research proposes a two-phase model for WDCNs including the service provision phase and the service access phase according to different traffic patterns and performance requirements. For the service provision phase, we investigate the impact of communication channel conditions on the average execution time of the computing tasks within WDCNs. We then discuses how to increase the robustness and power efficiency for WDCNs subject to the impact of channel variance and spatial heterogeneity. A resource allocation solution for computation oriented WDCNs is then introduced in detail which mitigates the effects of channel variations with a stochastic programming solution. Stochastic geometry and queue theory are combined to analyze the average performance of service response time and to design optimal access strategies during the service access phase. This access model provides a framework to analyze the service access performance and evaluate whether the channel heterogeneity should be considered. Based on this analysis, optimal strategies to access the service nodes can be determined in order to reduce the service response time. In addition, network initialization and synchronization are investigated in order to build a multiple channel WDCN in dynamic spectrum access (DSA) environments. Further, an efficient primary user detection method is proposed to reduce the channel vacation latency for WDCNs in DSA environments. Finally, this dissertation presents the complete design and implementation of a WDCN on COgnitive Radio Network (CORNET). Based on SDR technologies, software dedicated to WDCNs is designed and implemented across the PHY layer, MAC layer, and application layer. System experiments are carried out to demonstrate the performance issues and solutions presented in this dissertation. Wireless distributed computing networks (WDCNs) will become the next frontier of the wireless industry as the performance of wireless platforms is being increased every year and wireless industries are looking for "killer" applications for increased channel capacity. However, WDCNs have several unique problems compared with currently well-investigated methods for wireless sensor networks and wired distributed computing. For example, it is difficult for WDCNs to be power/energy efficient considering the uncertainty and heterogeneity of the wireless environment. In addition, the service model has to take account of the interference-limited feature of wireless channels to reduce the service delay. Our research proposes a two-phase model for WDCNs including the service provision phase and the service access phase according to different traffic patterns and performance requirements. For the service provision phase, we investigate the impact of communication channel conditions on the average execution time of the computing tasks within WDCNs. We then discuses how to increase the robustness and power efficiency for WDCNs subject to the impact of channel variance and spatial heterogeneity. A resource allocation solution for computation oriented WDCNs is then introduced in detail which mitigates the effects of channel variations with a stochastic programming solution. Stochastic geometry and queue theory are combined to analyze the average performance of service response time and to design optimal access strategies during the service access phase. This access model provides a framework to analyze the service access performance and evaluate whether the channel heterogeneity should be considered. Based on this analysis, optimal strategies to access the service nodes can be determined in order to reduce the service response time. In addition, network initialization and synchronization are investigated in order to build a multiple channel WDCN in dynamic spectrum access (DSA) environments. Further, an efficient primary user detection method is proposed to reduce the channel vacation latency for WDCNs in DSA environments. Finally, this dissertation presents the complete design and implementation of a WDCN on COgnitive Radio Network (CORNET). Based on SDR technologies, software dedicated to WDCNs is designed and implemented across the PHY layer, MAC layer, and application layer. System experiments are carried out to demonstrate the performance issues and solutions presented in this dissertation. / Ph. D.