No description available.

Due to the massive computational power, Graphics Processing Units (GPUs) have become a popular platform for executing general purpose parallel applications. The majority of on-chip communications in GPU architecture occur between memory controllers and compute cores, thus memory controllers become hot spots and bottle neck when conventional mesh interconnection networks are used. Leveraging this observation, we reduce the network latency and improve throughput by providing a nanophotonic ring network which connects all memory controllers. This new interconnection network employs a new routing algorithm that combines Dimension Ordered Routing (DOR) and nanophotonic ring algorithms. By exploring this new topology, we can achieve to reduce interconnection network latency by 17% on average (up to 32%) and improve IPC by 5% on average (up to 11.5%). We also analyze application characteristics of six CUDA benchmarks on the GPGPU-Sim simulator to obtain better perspective for designing high performance GPU interconnection network.

Master of Science / Department of Electrical and Computer Engineering / Caterina M. Scoglio / Recently, virtualization was proposed in many scientific fields. Virtualization is widely applied in telecommunications where networks are required to be extremely flexible to meet the current and the unpredictable future requirements. The creation of a virtual network over the physical network allows the application developers to design new services provided to the users without modifying the underlay resources. The creation of a virtual network of light paths and light trees over the optical network allows the resources managers to utilize the huge optical capacity more efficiently. In this thesis, we consider the optimal topology design for the virtual networks taking into consideration traffic demands and quality of service constraints of the applications. Considered examples of virtual networks are the overlay networks at the application layer and the virtual light path and light tree networks at the optical layer. In the design of overlay topologies, the performance of the virtual networks is affected by traffic characteristic, and behavior of nodes which can be selfish or cooperative. Both the static and dynamic traffic demand scenarios are considered. The static demand scenario follows well known probability distributions, while in the dynamic traffic scenario, the traffic matrix is predicted through measurements over each link in the network. We study the problem of finding the overlay topology that minimizes a cost function which takes into account the overlay link creation cost and the routing cost. We formulate the problem as an Integer Linear Programming and propose heuristics to find near-optimal overlay topologies with a reduced complexity. Virtual optical networks are designed to support many applications. Multicast sessions are an example of the applications running over the optical network. The main objective in creating the hybrid topology, composed by light paths and light trees, is to increase number of supported multicast sessions through sharing the network resources. The problem of establishing the hybrid topology is formulated using the Integer Linear Programming. Extensive data results and analysis are performed on the generated hybrid topologies for evaluation.

With the advent of distributed energy resources (DERs), electric vehicles, and demand-response programs, grid operators are in dire need of new monitoring and design tools that help improve efficiency, reliability, and stability of modern power networks. To this end, the work in this thesis explores a generalized modeling and analysis framework for two pertinent tasks: i) learning loads via grid probing, and; ii) optimizing power grid topologies for stability. Distribution grids currently lack comprehensive real-time metering. Nevertheless, grid operators require precise knowledge of loads and renewable generation to accomplish any feeder optimization task. At the same time, new grid technologies, such as solar panels and energy storage units are interfaced via inverters with advanced sensing and actuation capabilities. In this context, we first put forth the idea of engaging power electronics to probe an electric grid and record its voltage response at actuated and metered buses to infer non-metered loads. Probing can be accomplished by commanding inverters to momentarily perturb their power injections. Multiple probing actions can be induced within a few tens of seconds. Load inference via grid probing is formulated as an implicit nonlinear system identification task, which is shown to be topologically observable under certain conditions. The analysis holds for single- and multi-phase grids, radial or meshed, and applies to phasor or magnitude-only voltage data. Using probing to learn non-constant-power loads is also analyzed as a special case. Once a probing setup is deemed topologically observable, a methodology for designing probing injections abiding by inverter and network constraints to improve load estimates is provided. The probing task under noisy phasor and non-phasor data is tackled using a semidefinite-program relaxation. As a second contribution, we also study the effect of topology on the linear time-invariant dynamics of power networks. For a variety of stability metrics, a unified framework based on the H2-norm of the system is presented. The proposed framework assesses the robustness of power grids to small disturbances and is used to study the optimal placement of new lines on existing networks as well as the design of radial topologies for new networks. / Doctor of Philosophy / Increased penetration of distributed energy resources such as solar panels, wind farms, and energy storage systems is forcing utilities to rethink how they design and operate their power networks. To ensure efficient and reliable operation of distribution networks and to perform any grid-wide optimization or dispatch tasks, the system operator needs to precisely know the net load (energy output) of every customer. However, due to the sheer extent of distribution networks (millions of customers) and low investment interest in the past, distribution grids have limited metering infrastructure. Nevertheless, data from grid sensors comprised of voltage and load measurements are readily available from a subset of customers at high temporal resolution. In addition, the smart inverters found in solar panels, energy storage units, and electric vehicles can be controlled within microseconds. The work in this thesis explores how the proliferation of grid sensors together with the controllability of smart inverters can be leveraged for inferring the non-metered loads i.e., energy output of customers that are not equipped with smart inverters/sensors. In addition to the load learning task, this thesis also presents a modeling and analysis framework to study the optimal design of topologies (how customers are electrically inter-connected) for improving stability of our power networks.

A paradigm shift is underway in which the classical materials selection approach in engineering design is being replaced by the design of material structure and processing paths on a hierarchy of length scales for specific multifunctional performance requirements. In this dissertation, the focus is on designing mesoscopic material and product topology?? geometric arrangement of solid phases and voids on length scales larger than microstructures but smaller than the characteristic dimensions of an overall product. Increasingly, manufacturing, rapid prototyping, and materials processing techniques facilitate tailoring topology with high levels of detail. Fully leveraging these capabilities requires not only computational models but also a systematic, efficient design method for exploring, refining, and evaluating product and material topology and other design parameters for targeted multifunctional performance that is robust with respect to potential manufacturing, design, and operating variations. In this dissertation, the Robust Topological Preliminary Design Exploration Method is presented for designing complex multi-scale products and materials by topologically and parametrically tailoring them for multifunctional performance that is superior to that of standard designs and less sensitive to variations. A comprehensive robust design method is established for topology design applications. It includes computational techniques, guidelines, and a multiobjective decision formulation for evaluating and minimizing the impact of topological and parametric variation on the performance of a preliminary topological design. A method is also established for multifunctional topology design, including thermal topology design techniques and multi-stage, distributed design methods for designing preliminary topologies with built-in flexibility for subsequent modification for enhanced performance in secondary functional domains. Key aspects of the approach are demonstrated by designing linear cellular alloys??ered metallic cellular materials with extended prismatic cells?? three applications. Heat exchangers are designed with increased heat dissipation and structural load bearing capabilities relative to conventional heat sinks for microprocessor applications. Cellular materials are designed with structural properties that are robust to dimensional and topological imperfections such as missing cell walls. Finally, combustor liners are designed to increase operating temperatures and efficiencies and reduce harmful emissions for next-generation turbine engines via active cooling and load bearing within topologically and parametrically customized cellular materials.

Additive manufacturing technologies are enabling newfound degrees of geometrical complexity to be realised, particularly with regards to internal structures. All of these manufacturing technologies are dependant on their prior design in an appropriate electronic form, either by reverse engineering, or, primarily, by computer-aided design. Within these emerging applications is the design of scaffolds with an intricate and controlled internal structure for bone tissue engineering. There is a consensus that ideal bone scaffold geometry is evident in biological trabecular structures. In their most basic topological form,these structures consist of the non-linear distribution of irregular interconnecting rods and plates of different size and shape. Complex and irregular architectures can be realised by several scaffold manufacturing techniques, but with little or no control over the main features of the internal geometry, such as size, shape and interconnectivity of each individual element. The combined use of computer aided design systems and additive manufacturing techniques allows a high degree of control over these parameters with few limitations in terms of achievable complexity. However, the design of irregular and intricate trabecular networks in computer aided design systems is extremely time-consuming since manually modelling an extraordinary number of different rods and plates, all with different parameters, may require several days to design an individual scaffold structure. In an attempt to address these difficulties, several other research efforts in this domain have largely focussed on techniques which result in designs which comprise of relatively regular and primitive shapes and do not represent the level of complexity seen biologically. Detailed descriptions of these methods are covered in chapter 1. An automated design methodology for trabecular structures is proposed by this research to overcome these limitations. This approach involves the investigation of novel software algorithms, which are able to interact with a conventional computer aided design program and permit the automated design of geometrical elements in the form of rods, each with a different size and shape. The methodology is described in chapter 2 and is tested in chapter 3. Applications of this methodology in anatomical designs are covered in chapter 4. Nevertheless, complex designed rod networks may still present very different properties compared to trabecular bone geometries due to a lack detailed information available which explicitly detail their geometry. The lack of detailed quantitative descriptions of trabecular bone geometries may compromise the validity of any design methodology, irrespective of automation and efficiency. Although flexibility of a design methodology is beneficial, this may be rendered inadequate when insufficient quantitative data is known of the target structure. In this work a novel analysis methodology is proposed in chapter 5, which may provide a significant contribution toward the characterisation and quantification of target geometries, with particular focus on trabecular bone structures. This analysis methodology can be used either to evaluate existing design techniques or to drive the development of new bio-mimetic design techniques. This work then progresses to a newly derived bio-mimetic automated design technique, driven by the newly produced quantitative data on trabecular bone geometries. This advanced design methodology has been developed and tested in chapter 6. This has demonstrated the validity of the technique and realised a significant stage of development in the context and scope of this work.