No description available.

The shapes of cluster central galaxies are not randomly oriented, but rather exhibit coherent alignments with the shapes of their parent clusters as well as with the surrounding large-scale structures. In this work, we aim to identify the galaxy and cluster quantities that most strongly predict the central galaxy alignment phenomenon among a large parameter space with a sample of 8237 clusters and 94 817 members within 0.1 < z < 0.35, based on the red-sequence Matched-filter Probabilistic Percolation cluster catalogue constructed from the Sloan Digital Sky Survey. We first quantify the alignment between the projected central galaxy shapes and the distribution of member satellites, to understand what central galaxy and cluster properties most strongly correlate with these alignments. Next, we investigate the angular segregation of satellites with respect to their central galaxy major axis directions, to identify the satellite properties that most strongly predict their angular segregation. We find that central galaxies are more aligned with their member galaxy distributions in clusters that are more elongated and have higher richness, and for central galaxies with larger physical size, higher luminosity and centring probability, and redder colour. Satellites with redder colour, higher luminosity, located closer to the central galaxy, and with smaller ellipticity show a stronger angular segregation towards their central galaxy major axes. Finally, we provide physical explanations for some of the identified correlations, and discuss the connection to theories of central galaxy alignments, the impact of primordial alignments with tidal fields, and the importance of anisotropic accretion.

Large scale machine learning has many characteristics that can be exploited in the system designs to improve its efficiency. This dissertation demonstrates that the characteristics of the ML computations can be exploited in the design and implementation of parameter server systems, to greatly improve the efficiency by an order of magnitude or more. We support this thesis statement with three case study systems, IterStore, GeePS, and MLtuner. IterStore is an optimized parameter server system design that exploits the repeated data access pattern characteristic of ML computations. The designed optimizations allow IterStore to reduce the total run time of our ML benchmarks by up to 50×. GeePS is a parameter server that is specialized for deep learning on distributed GPUs. By exploiting the layer-by-layer data access and computation pattern of deep learning, GeePS provides almost linear scalability from single-machine baselines (13× more training throughput with 16 machines), and also supports neural networks that do not fit in GPU memory. MLtuner is a system for automatically tuning the training tunables of ML tasks. It exploits the characteristic that the best tunable settings can often be decided quickly with just a short trial time. By making use of optimization-guided online trial-and-error, MLtuner can robustly find and re-tune tunable settings for a variety of machine learning applications, including image classification, video classification, and matrix factorization, and is over an order of magnitude faster than traditional hyperparameter tuning approaches.

by Li Wai Ting. / Thesis (M.Ph.)--Chinese University of Hong Kong, 1987. / Includes bibliographical references.

Large-scale distributed systems and distributed computing are the pillars of IT infrastructure and society nowadays. Robust theoretical principles for designing, building, managing and understanding the interactive behaviours of such systems need to be explored. A promising approach for establishing such principles is to view the session as the key unit for design, execution and verification. Governance is a general term for verifying whether activities meet the specified requirements and for enforcing safe behaviours among processes. This thesis, based on the asynchronous -calculus and the theory of session types, provides a monitoring framework and a theory for validating specifications, verifying mutual behaviours during runtime, and taking actions when noncompliant behaviours are detected. We explore properties and principles for governing large-scale distributed systems, in which autonomous and heterogeneous system components interact with each other in the network to accomplish application goals. This thesis, incorporating lessons from my participation in a substantial practical project, the Ocean Observatories Initiative (OOI), proposes an asynchronous monitoring framework and the process calculus for dynamically governing the asynchronous interactions among distributed multiple applications. We prove that this monitoring model guarantees the satisfaction of global assertions, and state and prove theorems of local and global safety, transparency, and session fidelity. We also study and introduce the semantic mechanisms for runtime session-based governance and the principles of validation of stateful specifications through capturing the runtime asynchronous interactions.

He, Xu. / Thesis (Ph.D.)--Chinese University of Hong Kong, 2013. / Includes bibliographical references (leaves [127]-135). / Electronic reproduction. Hong Kong : Chinese University of Hong Kong, [2012] System requirements: Adobe Acrobat Reader. Available via World Wide Web. / Abstracts also in Chinese.

Observational surveys which probe our universe deeper and deeper into the nonlinear regime of structure formation are becoming increasing accurate. This makes numerical simulations an essential tool for theory to be able to predict phenomena at comparable scales. In the first part of this thesis we study the behaviour of cosmological models involving a scalar field. We are particularly interested in the existence of fixed points of the dynamical system and the behaviour of the system in their vicinity. Upon addition of spatial curvature to the single-scalar field model with an exponential potential, canonical kinetic term, and a matter fluid, we demonstrate the existence of two extra fixed points that are not present in the case without curvature. We also analyse the evolution of the equation-of-state parameter. In the second part, we numerically simulate collisionless particles in the weak field approximation to General Relativity, with large gradients of the fields and relativistic velocities allowed. To reduce the complexity of the problem and enable high resolution simulations, we consider the spherically symmetric case. Comparing numerical solutions to the exact Schwarzschild and Lemaître-Tolman-Bondi solutions, we show that the scheme we use is more accurate than a Newtonian scheme, correctly reproducing the leading-order post-Newtonian behaviour. Furthermore, by introducing angular momentum, configurations corresponding to bound objects are found. In the final part, we simulate the conditions under which one would expect to form ultracompact minihalos, dark matter halos with a steep power-law profile. We show that an isolated object exhibits the profile predicted analytically. Embedding this halo in a perturbed environment we show that its profile becomes progressively more similar to the Navarro-Frenk-White profile with increasing amplitude of perturbations. Next, we boost the power spectrum at a very early redshift during radiation domination on a chosen scale and simulate clustering of dark matter particles at this scale until low redshift. In this scenario halos form earlier, have higher central densities, and are more compact.

The ever-increasing design complexity of modern circuits challenges our ability to verify their correctness. Therefore, various errors are more likely to escape the pre-silicon verification process and to manifest themselves after design tape-out. To address this problem, effective post-silicon validation is essential for eliminating design bugs before integrated circuit (IC) products shipped to customers. In the debug process, it becomes increasingly popular to insert design-for-debug (DfD) structures into the original design to facilitate real-time debug without intervening the circuits’ normal operation. For this so-called trace-based post-silicon validation technique, the key question is how to design such DfD circuits to achieve sufficient observability and controllability during the debug process with limited hardware overhead. However, in today’s VLSI design flow, this is unfortunately conducted in a manual fashion based on designers’ own experience, which cannot guarantee debug quality. To tackle this problem, we propose a set of automatic tracing solutions as well as innovative DfD designs in this thesis. First, we develop a novel trace signal selection technique to maximize the visibility on debugging functional design errors. To strengthen the capability for tackling these errors, we sequentially introduce a multiplexed signal tracing strategy with a trace signal grouping algorithm for maximizing the probability of catching the propagated evidences from functional design errors. Then, to effectively localize speedpathrelated electrical errors, we propose an innovative trace signal selection solution as well as a trace qualification technique. On the other hand, we introduce several low-cost interconnection fabrics to effectively transfer trace data in post-silicon validation. We first propose to reuse the existing test channel for real-time trace data transfer, so that the routing cost of debug hardware is dramatically reduced. The method is further improved to avoid data corruption in multi-core debug. We then develop a novel interconnection fabric design and optimization technique, by combining multiplexor network and non-blocking network, to achieve high debug flexibility with minimized hardware cost. Moreover, we introduce a hybrid trace interconnection fabric that is able to tolerate unknown values in “golden vectors“, at the cost of little extra DfD overhead. With the fabric, we develop a systematic signal tracing procedure to automatically localize erroneous signals with just a few debug runs. Our empirical evaluation shows that the solutions presented in this thesis can greatly improve the validation quality of VLSI circuits, and ultimately enable the design and fabrication of reliable electronic devices. / Liu, Xiao. / Thesis (Ph.D.)--Chinese University of Hong Kong, 2012. / Includes bibliographical references (leaves 143-152). / Electronic reproduction. Hong Kong : Chinese University of Hong Kong, [2012] System requirements: Adobe Acrobat Reader. Available via World Wide Web. / Abstract --- p.i / Acknowledgement --- p.iv / Preface --- p.vii / Chapter 1 --- Introduction --- p.1 / Chapter 1.1 --- VLSI Design Trends and Validation Challenges --- p.1 / Chapter 1.2 --- Key Contributions and Thesis Outline --- p.4 / Chapter 2 --- State of the Art on Post-Silicon Validation --- p.8 / Chapter 2.1 --- Trace Signal Selection --- p.12 / Chapter 2.2 --- Interconnection Fabric Design for Trace Data Transfer --- p.14 / Chapter 2.3 --- Trace Data Compression --- p.15 / Chapter 2.4 --- Trace-Based Debug Control --- p.16 / Chapter 3 --- Signal Selection for Visibility Enhancement --- p.18 / Chapter 3.1 --- Preliminaries and Summary of Contributions --- p.19 / Chapter 3.2 --- Restorability Formulation --- p.23 / Chapter 3.2.1 --- Terminologies --- p.23 / Chapter 3.2.2 --- Gate-Level Restorabilities --- p.24 / Chapter 3.3 --- Trace Signal Selection --- p.28 / Chapter 3.3.1 --- Circuit Level Visibility Calculation --- p.28 / Chapter 3.3.2 --- Trace Signal Selection Methodology --- p.30 / Chapter 3.3.3 --- Trace Signal Selection Enhancements --- p.31 / Chapter 3.4 --- Experimental Results --- p.34 / Chapter 3.4.1 --- Experiment Setup --- p.34 / Chapter 3.4.2 --- Experimental Results --- p.35 / Chapter 3.5 --- Conclusion --- p.40 / Chapter 4 --- Multiplexed Tracing for Design Error --- p.47 / Chapter 4.1 --- Preliminaries and Summary of Contributions --- p.49 / Chapter 4.2 --- Design Error Visibility Metric --- p.53 / Chapter 4.3 --- Proposed Methodology --- p.56 / Chapter 4.3.1 --- Supporting DfD Hardware for Multiplexed Signal Tracing --- p.58 / Chapter 4.3.2 --- Signal Grouping Algorithm --- p.58 / Chapter 4.4 --- Experimental Results --- p.62 / Chapter 4.4.1 --- Experiment Setup --- p.62 / Chapter 4.4.2 --- Experimental Results --- p.63 / Chapter 4.5 --- Conclusion --- p.66 / Chapter 5 --- Tracing for Electrical Error --- p.68 / Chapter 5.1 --- Preliminaries and Summary of Contributions --- p.69 / Chapter 5.2 --- Observing Speedpath-Related Electrical Errors --- p.71 / Chapter 5.2.1 --- Speedpath-Related Electrical Error Model --- p.71 / Chapter 5.2.2 --- Speedpath-Related Electrical Error Detection Quality --- p.73 / Chapter 5.3 --- Trace Signal Selection --- p.75 / Chapter 5.3.1 --- Relation Cube Extraction --- p.76 / Chapter 5.3.2 --- Signal Selection for Non-Zero-Probability Error Detection --- p.77 / Chapter 5.3.3 --- Trace Signal Selection for Error Detection Quality Enhancement --- p.78 / Chapter 5.4 --- Trace Data Qualification --- p.80 / Chapter 5.5 --- Experimental Results --- p.83 / Chapter 5.6 --- Conclusion --- p.87 / Chapter 6 --- Reusing Test Access Mechanisms --- p.88 / Chapter 6.1 --- Preliminaries and Summary of Contributions --- p.89 / Chapter 6.1.1 --- SoC Test Architectures --- p.89 / Chapter 6.1.2 --- SoC Post-Silicon Validation Architectures --- p.90 / Chapter 6.1.3 --- Summary of Contributions --- p.92 / Chapter 6.2 --- Overview of the Proposed Debug Data Transfer Framework --- p.93 / Chapter 6.3 --- Proposed DfD Structures --- p.94 / Chapter 6.3.1 --- Modified Wrapper Design --- p.95 / Chapter 6.3.2 --- Trace Buffer Interface Design --- p.97 / Chapter 6.4 --- Sharing TAM for Multi-Core Debug Data Transfer --- p.98 / Chapter 6.4.1 --- Core Masking for TestRail Architecture --- p.98 / Chapter 6.4.2 --- Channel Split --- p.99 / Chapter 6.5 --- Experimental Results --- p.101 / Chapter 6.6 --- Conclusion --- p.104 / Chapter 7 --- Interconnection Fabric for Flexible Tracing --- p.105 / Chapter 7.1 --- Preliminaries and Summary of Contributions --- p.106 / Chapter 7.2 --- Proposed Interconnection Fabric Design --- p.111 / Chapter 7.2.1 --- Multiplexer Network for Mutually-Exclusive Signals --- p.111 / Chapter 7.2.2 --- Non-Blocking Concentration Network for Concurrently-Accessible Signals --- p.114 / Chapter 7.3 --- Experimental Results --- p.117 / Chapter 7.4 --- Conclusion --- p.121 / Chapter 8 --- Interconnection Fabric for Systematic Tracing --- p.123 / Chapter 8.1 --- Preliminaries and Summary of Contributions --- p.124 / Chapter 8.2 --- Proposed Trace Interconnection Fabric --- p.128 / Chapter 8.3 --- Proposed Error Evidence Localization Methodology --- p.130 / Chapter 8.4 --- Experimental Results --- p.133 / Chapter 8.4.1 --- Experimental Setup --- p.133 / Chapter 8.4.2 --- Results and Discussion --- p.134 / Chapter 8.5 --- Conclusion --- p.139 / Chapter 9 --- Conclusion --- p.140 / Bibliography --- p.152

In this work we explore three-point statistics applied to the large-scale structure in our Universe. Three-point statistics, such as the bispectrum, encode information not accessible via the standard analysis method-the power spectrum-and thus provide the potential for greatly improving current constraints on cosmological parameters. They also present us with additional challenges, and we focus on two of these arising from a measurement as well as modelling point of view. The first challenge we address is the covariance matrix of the bispectrum, as its precise estimate is required when performing likelihood analyses. Covariance matrices are usually estimated from a set of independent simulations, whose minimum number scales with the dimension of the covariance matrix. Because there are many more possibilities of finding triplets of galaxies than pairs, compared to the power spectrum this approach becomes rather prohibitive. With this motivation in mind, we explore a novel alternative to the bispectrum: the line correlation function (LCF). It specifically targets information in the phases of density modes that are invisible to the power spectrum, making it a potentially more efficient probe than the bispectrum, which measures a combination of amplitudes and phases. We derive the covariance properties and the impact of shot noise for the LCF and compare these theoretical predictions with measurements from N-body simulations. Based on a Fisher analysis we assess the LCF's sensitivity on cosmological parameters, finding that it is particularly suited for constraining galaxy bias parameters and the amplitude of fluctuations. As a next step we contrast the Fisher information of the LCF with the full bispectrum and two other recently proposed alternatives. We show that the LCF is unlikely to achieve a lossless compression of the bispectrum information, whereas a modal decomposition of the bispectrumcan reduce the size of the covariancematrix by at least an order of magnitude. The second challenge we consider in this work concerns the relation between the dark matter field and luminous tracers, such as galaxies. Accurate knowledge of this galaxy bias relation is required in order to reliably interpret the data gathered by galaxy surveys. On the largest scales the dark matter and galaxy densities are linearly related, but a variety of additional terms need to be taken into account when studying clustering on smaller scales. These have been fully included in recent power spectrumanalyses, whereas the bispectrummodel relied on simple prescriptions that were likely extended beyond their realm of validity. In addition, treating power spectrumand bispectrum on different footings means that the two models become inconsistent on small scales. We introduce a new formalism that allows us to elegantly compute the lacking bispectrum contributions from galaxy bias, without running into the renormalization problem. Furthermore, we fit our new model to simulated data by implementing these contributions into a likelihood code. We show that they are crucial in order to obtain results consistent with those fromthe power spectrum, and that the bispectrum retains its capability of significantly reducing uncertainties in measured parameters when combined with the power spectrum.

There has been an increased interest in optimization for the analysis of large-scale data sets which require gigabytes or terabytes of data to be stored. A variety of applications originate from the fields of signal processing, machine learning and statistics. Seven representative applications are described below. - Magnetic Resonance Imaging (MRI): A medical imaging tool used to scan the anatomy and the physiology of a body. - Image inpainting: A technique for reconstructing degraded parts of an image. - Image deblurring: Image processing tool for removing the blurriness of a photo caused by natural phenomena, such as motion. - Radar pulse reconstruction. - Genome-Wide Association study (GWA): DNA comparison between two groups of people (with/without a disease) in order to investigate factors that a disease depends on. - Recommendation systems: Classification of data (i.e., music or video) based on user preferences. - Data fitting: Sampled data are used to simulate the behaviour of observed quantities. For example estimation of global temperature based on historic data. Large-scale problems impose restrictions on methods that have been so far employed. The new methods have to be memory efficient and ideally, within seconds they should offer noticeable progress towards a solution. First-order methods meet some of these requirements. They avoid matrix factorizations, they have low memory requirements, additionally, they sometimes offer fast progress in the initial stages of optimization. Unfortunately, as demonstrated by numerical experiments in this thesis, first-order methods miss essential information about the conditioning of the problems, which might result in slow practical convergence. The main advantage of first-order methods which is to rely only on simple gradient or coordinate updates becomes their essential weakness. We do not think this inherent weakness of first-order methods can be remedied. For this reason, the present thesis aims at the development and implementation of inexpensive higher-order methods for large-scale problems.