Downloads & Free Reading Options - Results

Available from National Technical Information Service, Springfield, Va

Available from National Technical Information Service, Springfield, Va

There has been some additional development of parts reported in "Multi-Modulator for Bandwidth-Efficient Communication" (NPO-40807), NASA Tech Briefs, Vol. 32, No. 6 (June 2009), page 34. The focus was on 1) The generation of M-order quadrature amplitude modulation (M-QAM) and octonary-phase-shift-keying, trellis-coded modulation (8PSK TCM), 2) The use of square-root raised-cosine pulse-shaping filters, 3) A parallel-processing architecture that enables low-speed [complementary metal oxide/semiconductor (CMOS)] circuitry to perform the coding, modulation, and pulse-shaping computations at a high rate; and 4) Implementation of the architecture in a CMOS field-programmable gate array.

A 1-bit serial processor designed for a parallel computer architecture is described. This processor is used to develop a massively parallel computational engine, with a single instruction-multiple data (SIMD) architecture. The computer is simulated and tested to verify its operation and to measure its performance for further development.

Many computing tasks involve heavy mathematical calculations, or analyzing large amounts of data. These operations can take a long time to complete using only one computer. Networks such as the Internet provide many computers with the ability to communicate with each other. Parallel or distributed computing takes advantage of these networked computers by arranging them to work together on a problem, thereby reducing the time needed to obtain the solution. The drawback to using a network of computers to solve a problem is the time wasted in communicating between the various hosts. The application of distributed computing techniques to a space environment or to use over a satellite network would therefore be limited by the amount of time needed to send data across the network, which would typically take much longer than on a terrestrial network. This experiment shows how much faster a large job can be performed by adding more computers to the task, what role communications time plays in the total execution time, and the impact a long-delay network has on a distributed computing system.

Host plus multiple array processor architecture is demonstrated to yield a modular, fast, and cost-effective system for radar processing. Software methodology for programming such a system is developed. Parallel processing with pipelined data flow among the host, array processors, and discs is implemented. Theoretical analysis of performance is made and experimentally verified. The broad class of problems to which the architecture and methodology can be applied is indicated.

The principal objective of this research is to develop, test, and implement coarse-grained, parallel-processing strategies for nonlinear dynamic simulations of practical structural problems. There are contributions to four main areas: finite element modeling and analysis of rotational dynamics, numerical algorithms for parallel nonlinear solutions, automatic partitioning techniques to effect load-balancing among processors, and an integrated parallel analysis system.

The principal objective of this research is to develop, test, and implement coarse-grained, parallel-processing strategies for nonlinear dynamic simulations of practical structural problems. There are contributions to four main areas: finite element modeling and analysis of rotational dynamics, numerical algorithms for parallel nonlinear solutions, automatic partitioning techniques to effect load-balancing among processors, and an integrated parallel analysis system.

The high resolution direction-of-arrival (DOA) estimation is important in many sensor systems. It is based on the processing of the received signal and extracting the desired parameters of the DOA of plane waves. Many approaches have been used for the purpose of implementing the function required for the DOA estimation. The Multiple Signal Classification (MUSIC) and the Estimation of Signal Parameters by Rotational Invariance techniques (ESPRIT) algorithms are two novel approaches used recently to provide asymptotically unbiased and efficient estimates of the DOA. They are believed to be promising and appropriate for hardware implementation for real time applications. They estimate the so called signal subspace from the array measurements. The parameters of interest (i.e. determining of the DOA) are then estimated from the intersection between the array manifold and the estimated subspace

The present study shows that search for a conjunction of color and orientation can be performed in parallel when the target is embedded in a homogeneous colored subgroup of nontarget elements. The results provide evidence for the notion of sequential global-to-local parallel processing in which the first global stage allows the rejection of one subgroup of elements followed by a second local stage which rejects the local subgroup of elements. Experiment 1 shows that parallel search for a conjunction target can only occur when the elements are spatially grouped. Experiment 2 shows that parallel processing of a conjunction target is not limited to small display sizes but also occur in displays consisting of 49 elements. Implications of the present findings for current theories of visual search are discussed.

The Final Proceedings for VECPAR, 4th International Meeting on Vector and Parallel Processing, 21-23 June 2000, an interdisciplinary conference covering topics in all areas of vector, parallel and distributed computing applied to a broad range of research disciplines with a focus on engineering. The principal topics include: Cellular Automata, Computational Fluid Dynamics, Crash and Structural Analysis, Data Warehousing and Data Mining, Distributed Computing and Operating Systems, Fault Tolerant Systems, Imaging and Graphics, Interconnection Networks, Languages and Tools, Numerical Methods, Parallel and Distributed Algorithms, Real-time and Embedded Systems, Reconfigurable Systems, Linear Algebra Algorithms and Software for Large Scientific Problems, Computer Organization, Image Analysis and Synthesis, and Nonlinear Problems.

The purpose of this study was to evaluate the parallel processing impact of MIMD (Multiple-Instruction Multiple-Data Path) computers on Flight Simulation Software. Basic mathematical functions and arithmetic expressions from typical Flight Simulation Software were selected and run on a MIMD computer to evaluate the improvement in execution time that results from the parallel architecture of this type of computer. Recommendations as to the types of tasks which are optimally suitable for this computer architecture are made, together with the improvement in execution speed to be expected. (Author)

A large rule-based expert system with each rule involving perhaps 10 out of l00,000 possible Boolean criteria, can require a significant amount of processing time to evaluate. This time can be reduced if all rules have a single consequent and have antecedents that contain only conjunctions of the Boolean criteria or their complements. If the consequences do not insert new facts into the rule-base, then parallel processing can be used with great efficiency. The value of a rule-based expert system to help solve a variety of diagnostic and advisory needs has been well-demonstrated over the last 2 decades. Parallel processing has become increasingly important for embedded systems in order to accelerate a variety of computations. This report discusses research connected to the development of a data structure and algorithm to perform parallel inferencing in rule-based systems. It also discusses a simulation technique for estimating the number of processors needed to evaluate a given number of rules and criteria within the required time. Expert systems, Knowledge-based systems, Rule-based Systems, Artificial intelligence, Real-time Processing, Parallel processing, Distributed computing, Decision support, Cockpit Automation, Function allocation, Simulation.

Molecular dynamics (MD) implemented on parallel processors was used to model supercritical droplet phenomena occurring in combustion devices. The use of molecular dynamics allows the modeling of supercritical phenomena without an a priori knowledge of the equation of state or transport properties of the individual components or the mixture. Three-dimensional supercritical oxygen vaporization into gaseous oxygen and helium using two-site Lennard-Jones potentials for the oxygen has been modeled and both the disappearance of surface tension above the critical point and the modification of the critical point for a binary mixture have been observed. A distinct change in droplet morphology was observed when passing through its critical point. The droplet remains spherical as it vaporizes under subcritical conditions but becomes broken and cloud-like when supercritical. Equations of state and transport coefficients for mass, momentum and energy have been calculated for supercritical argon, nitrogen and oxygen which agree with NIST values.

The numerical solution of many problems in continuum dynamics is seriously limited by the computation rates attainable on computers with serial architecture. Parallel processing machines can achieve much higher rates. However, applying additional processors to a calculation is only part of the solution. This research was undertaken to develop parallel algorithms for explicit and implicit, Lagrangian and Eulerian finite difference schemes for computational continuum dynamics in one spatial dimension. First, the explicit conservation equations in the Lagrangian reference frame were readily reformulated for concurrent processing. Second, an implicit solution was derived for these equations. This is important because it yields unconditional stability. The parallelism is achieved bia a block implicit numerical scheme. Third, a rezoning algorithm was employed with each Lagrangian integration stem to transform the mesh back to the Eulerian reference frame. The algorithmic development path lead to a parallelization of the processing in blocks of the finite difference zones. AWt each step of this research project, the derived numerical methods provided effective algorithms for exploiting the architectural advantages of the HEP H1000 (Heterogeneous Element Processor) computer. The computational timing data shows significant speed-up with the number of processes. (Author)

The objective of this research is to provide a design methodology for connecting heterogeneous hardware modules that have inherently different functional and timing behavior. With the constraints dictated by the system- level interaction, we need to adopt a modular design approach without compromising the global performance. The main task of this effort will be the development of the theory for optimal interface circuits synthesis from high- level specification, with emphasis on testability and performance. In July-Dec 1991 we have concentrated on the gate-level synthesis of asynchronous control circuits and introducing timing information into the synthesis process to increase circuit performance.

This Memorandum describes work done in mapping image processing algorithms on to transputer arrays. Object classification based on the autocorrelation function has been applied to infrared images. Using the OCCAM programming language, parallel processing techniques have been developed to make use of the concurrency available for communication and calculation. The results for this specific problem are assessed and conclusions drawn on the use of transputers and OCCAM for parallel processing.

This Memorandum describes work done in mapping image processing algorithms on to transputer arrays. Object classification based on the autocorrelation function has been applied to infrared images. Using the OCCAM programming language, parallel processing techniques have been developed to make use of the concurrency available for communication and calculation. The results for this specific problem are assessed and conclusions drawn on the use of transputers and OCCAM for parallel processing.

This report presents the results of research on the allocation of binary cartesian product database files across several stores (such as disks) which are accessible in parallel so that the average time required to perform the retrieval (by using several parallel accesses) of all buckets specified by a partial-match query is objective of constructing efficient packings of Hamming spheres in hypercubes. The viewpoint of binary cartesian product file allocation thus produced allowed in our research the creation of allocations for two particular sizes of the set of stores such that the average processing time for partial-match queries was smaller than that of the best allocation technique found in the literature for these same two cases. It also provides a framework in which good allocations for other particular sizes of the set of stores, and possibly for general sets of stores, might be found.

This report presents the results of research on the allocation of binary cartesian product database files across several stores (such as disks) which are accessible in parallel so that the average time required to perform the retrieval (by using several parallel accesses) of all buckets specified by a partial-match query is objective of constructing efficient packings of Hamming spheres in hypercubes. The viewpoint of binary cartesian product file allocation thus produced allowed in our research the creation of allocations for two particular sizes of the set of stores such that the average processing time for partial-match queries was smaller than that of the best allocation technique found in the literature for these same two cases. It also provides a framework in which good allocations for other particular sizes of the set of stores, and possibly for general sets of stores, might be found.

Large industrial systems that combine services and applications, have become targets for cyber criminals and are challenging from the security, monitoring and auditing perspectives. Security log analysis is a key step for uncovering anomalies, detecting intrusion, and enabling incident response. The constant increase of link speeds, threats and users, produce large volumes of log data and become increasingly difficult to analyse on a Central Processing Unit (CPU). This paper presents a massively parallel Graphics Processing Unit (GPU) LOg Processing (GLoP) library and can also be used for Deep Packet Inspection (DPI), using a prefix matching technique, harvesting the full power of off-the-shelf technologies. GLoP implements two different algorithm using different GPU memory and is compared against CPU counterpart implementations. The library can be used for processing nodes with single or multiple GPUs as well as GPU cloud farms. The results show throughput of 20Gbps and demonstrate that modern GPUs can be utilised to increase the operational speed of large scale log processing scenarios, saving precious time before and after an intrusion has occurred.

Large industrial systems that combine services and applications, have become targets for cyber criminals and are challenging from the security, monitoring and auditing perspectives. Security log analysis is a key step for uncovering anomalies, detecting intrusion, and enabling incident response. The constant increase of link speeds, threats and users, produce large volumes of log data and become increasingly difficult to analyse on a Central Processing Unit (CPU). This paper presents a massively parallel Graphics Processing Unit (GPU) LOg Processing (GLoP) library and can also be used for Deep Packet Inspection (DPI), using a prefix matching technique, harvesting the full power of off-the-shelf technologies. GLoP implements two different algorithm using different GPU memory and is compared against CPU counterpart implementations. The library can be used for processing nodes with single or multiple GPUs as well as GPU cloud farms. The results show throughput of 20Gbps and demonstrate that modern GPUs can be utilised to increase the operational speed of large scale log processing scenarios, saving precious time before and after an intrusion has occurred.

The intent of this thesis is to answer the question as to whether real time battlefield visualization, once requiring high speed UNIX workstations and specialized parallel processors, can now be now performed on relatively inexpensive off the shelf components. Alternative network topologies were implemented using 10 and 100 megabit-per-second Ethernet cards under the Linux operating system on Pentium based personal computer platforms. Network throughput, processor and video performance benchmark routines were developed to assess the hardware's potential for parallel application in a distributed environment. Code was first ported to the Linux environment. Benchmark routines were then developed and tested on various machines. Dual 200 MHz Pentium Pro processor performance exceeded the dual processor 50 MHz SUN and 40 MHz SGI UNIX workstations currently used for terrain generation by a factor of 30 using a simple ray trace algorithms as a basis for comparison. The Intel Pentium Pro personal computer proved to be a capable platform for generating six to ten frame-per-second terrain simulations. However, Fast Ethernet throughput averages only 2.5 megabytes-per-second, thereby limiting the usefulness of a distributed approach designed to increase performance by dividing workload across the network.

Many scientific data-intensive applications perform iterative computations on array data. There exist multiple engines specialized for array processing. These engines efficiently support various types of operations, but none includes native support for iterative processing. In this paper, we develop a model for iterative array computations and a series of optimizations. We evaluate the benefits of an optimized, native support for iterative array processing on the SciDB engine and real workloads from the astronomy domain.

http://uf.catalog.fcla.edu/uf.jsp?st=UF002097523&ix=pm&I=0&V=D&pm=1

http://uf.catalog.fcla.edu/uf.jsp?st=UF002097523&ix=pm&I=0&V=D&pm=1

Processing moving object trajectories arises in many application domains and has been addressed by practitioners in the spatiotemporal database and Geographical Information System communities. In this work, we focus on a trajectory similarity search, the distance threshold query, which finds all trajectories within a given distance d of a search trajectory over a time interval. We demonstrate the performance of a multithreaded implementation which features the use of an R-tree index and which has high parallel efficiency (78%-90%). We introduce a GPGPU implementation which avoids the use of index-trees, and instead features a GPU-friendly indexing method. We compare the performance of the multithreaded and GPU implementations, and show that a speedup can be obtained using the latter. We propose two classes of algorithms, SetSplit and GreedySetSplit, to create efficient query batches that reduce memory pressure and computational cost on the GPU. However, we find that using fixed-size batches is sufficiently efficient in practice. We develop an empirical performance model for our GPGPU implementation that can be used to predict the response time of the distance threshold query. This model can be used to pick a good query batch size.

This article is from Sensors (Basel, Switzerland) , volume 9 . Abstract In this study, a parallel processing method using a PC cluster and a virtual grid is proposed for the fast processing of enormous amounts of airborne laser scanning (ALS) data. The method creates a raster digital surface model (DSM) by interpolating point data with inverse distance weighting (IDW), and produces a digital terrain model (DTM) by local minimum filtering of the DSM. To make a consistent comparison of performance between sequential and parallel processing approaches, the means of dealing with boundary data and of selecting interpolation centers were controlled for each processing node in parallel approach. To test the speedup, efficiency and linearity of the proposed algorithm, actual ALS data up to 134 million points were processed with a PC cluster consisting of one master node and eight slave nodes. The results showed that parallel processing provides better performance when the computational overhead, the number of processors, and the data size become large. It was verified that the proposed algorithm is a linear time operation and that the products obtained by parallel processing are identical to those produced by sequential processing.

This article is from Sensors (Basel, Switzerland) , volume 9 . Abstract In this study, a parallel processing method using a PC cluster and a virtual grid is proposed for the fast processing of enormous amounts of airborne laser scanning (ALS) data. The method creates a raster digital surface model (DSM) by interpolating point data with inverse distance weighting (IDW), and produces a digital terrain model (DTM) by local minimum filtering of the DSM. To make a consistent comparison of performance between sequential and parallel processing approaches, the means of dealing with boundary data and of selecting interpolation centers were controlled for each processing node in parallel approach. To test the speedup, efficiency and linearity of the proposed algorithm, actual ALS data up to 134 million points were processed with a PC cluster consisting of one master node and eight slave nodes. The results showed that parallel processing provides better performance when the computational overhead, the number of processors, and the data size become large. It was verified that the proposed algorithm is a linear time operation and that the products obtained by parallel processing are identical to those produced by sequential processing.

This article is from Bioinformatics , volume 30 . Abstract Motivation: Using high-throughput sequencing, researchers are now generating hundreds of whole-genome assays to measure various features such as transcription factor binding, histone marks, DNA methylation or RNA transcription. Displaying so much data generally leads to a confusing accumulation of plots. We describe here a multithreaded library that computes statistics on large numbers of datasets (Wiggle, BigWig, Bed, BigBed and BAM), generating statistical summaries within minutes with limited memory requirements, whether on the whole genome or on selected regions.Availability and Implementation: The code is freely available under Apache 2.0 license at www.github.com/Ensembl/WiggletoolsContact:[email protected] or [email protected]

From the bitsavers.org collection, a scanned-in computer-related document. cdc :: cyber :: cyber 170 180 Models Hardware Ref Mans :: Cyberplus :: 60461830H Cyberplus Parallel Processing System Vol3 Diagrams 198607

This project consists of two parts: applications of wavelet transforms in multiscale image processing, and parallel algorithms and architectures. We have studied issues in wavelet-based edge detection: antisymmetry of wavelet filters and their support size with respect to edge localization, and a non-orthogonal four-coefficient wavelet edge detector. Texture segmentation using a modulated Daubechies wavelet has been studied, providing both spatial frequency and orientation selectivity. Object segmentation has been studied by examining silhouettes at multiple resolutions in piecewise linear approximations, and has been explored on LADAR data for use in target recognition. Applications to biomedical image compression, image halftoning and artificial neural network structure have also been investigated. In parallel processing, we have studied embeddings of wavelet transform algorithms as well as other algorithms for 2D and 3D mesh architectures and systolic arrays, and their time complexities have been evaluated. A variety of issues are addressed in utilization of 3D meshes in parallel image processing on 2D and 3D images including: the embedding of 2D images into 3D meshes, 3D shrinking incorporating subfields methodology, 3D connected component labeling, and segmentation in magnetic resonance imaging. Some fundamental issues for parallel reduction and reduction-augmentation operators in 2D and 3D image spaces have been examined.

The software development process for parallel processors is investigated with a focus on real-time sensor processing applications. Concepts that are relevant to real-time parallel processing are introduced including: a definition of scalable dataflow graphs motivated by the need to meet a fixed throughput constraint for varying problem sizes, an algorithm classification that makes explicit the impact that data dependencies have in real-time implementations, and a real-time implementation strategy that decomposes the most problematic algorithms into compositions of more predictable constituents and then uses scalable dataflow graphs and parallel processing to recover timing predictability by mapping data dependent timing uncertainties into the spatial dimension (processors). Two case studies apply these ideas: an implementation of the Modified Gram-Schmidt (MGS) algorithm on a MasPar MP1 and an implementation of the joint probabilistic data association (JPDA) algorithm on a Thinking Machines CM-2. The JPDA case study includes a SISAL implementation to illustrate the advantages of functional programming for these applications. Parallel processors, Scalable dataflow graphs, Algorithms.

Simulations: Processing of chaos and memory storage. In view of our previously published findings showing that motor patterns represent adaptive behaviors may be generated by chaotic activity, we have used computer simulations to examine the ability of simple networks to learn to process chaotic signals and to perform complex operations on them. These studies have shown that even simple networks can be used to understand how networks store information, much of which information can not have been obtained from the more complex biological systems. As one example, an important and unexpected finding is that networks having trainable thresholds, in addition to trainable synapses, can performs computations that trainable synapses alone can not, regardless of the number synapses that may be included in the network. Another finding is that when networks must learn several tasks simultaneously, the effective size network is self-limiting, and probably does not require special algorithmic rules for limiting the size of successfully computing neural connections.

Direct numerical simulations of separated-reattaching and separated flows have been performed on massively parallel processing computers. Two basic geometrical configurations have been studied: the separated-reattaching flow past a normal flat plate with an attached downstream splitter plate and the separated flow past a flat plate held normal to a uniform stream. A high-order finite-difference formulation on collocated grids has been developed to perform unsteady fluid flow simulations in rectangular geometries. The numerical procedure is based on a fifth-order upwind-biased scheme for the convective terms and a fourth-order accurate stencil for the diffusive terms. A direct solver based on eigenvalue decomposition has been developed for the pressure-Poisson equation. A mixed Fourier spectral /finite difference formulation is used for the spanwise discretization, and a data-parallel algorithm has been developed for the CM-5. The performance of the algorithm has been evaluated on various grid sizes in model flow problems and for different partition sizes. The characteristics of the separated-reattaching flow have been investigated through two-dimensional simulations in the steady and unsteady regimes. The shedding mechanism is characterized by two major modes at Re = 250 and a single mode at Re = 375 and 500. Further, the instability of the separated shear layer is found to be consistent with inviscid theory. For the two-dimensional study of the separated flow past a normal flat plate, the time-mean flow quantities are observed to be over-estimated compared to the experiments. The time-mean drag coefficient is also over-predicted by a factor of up to 2. This is attributed to the high coherence of the vortices predicted by the two-dimensional simulations

Comparison of biological (DNA or protein) sequences provides insight into molecular structure, function, and homology, and is increasingly important as the available data bases become larger and more numerous. One method of increasing the speed of the calculations is to perform them in parallel. We present the results of initial investigations using the Intel iPSC hypercube and the Connection Machine for these comparisons. Since these machines have very different architectures, the issues and performance trade-offs discussed have a wide applicability for the parallel processing of biological sequence comparisons.

This research was conducted to obtain better methods for image processing. It focused on several aspects of this problem, including parallel algorithms for image processing, knowledge-based techniques for image understanding, and modeling images using shape and texture. Eighteen technical reports were produced which will also appear as published papers in journals. In the paper Holes and Genus of 3D images, it was shown that certain geometric invariants of a digital image (number of components, number of holes, and number of cavities) do not determine the topology (in the sense of connectivity) of the image refuting the commonly believed assumption that they do. This research lays the groundwork for research on digital and computational geometry of 3D images. In the paper Hough Transform Algorithms for Mesh-Connected for Mesh-Connected SIMD Parallel Processors, several methods of Hough transform computation are studied in terms of suitability for implementation on a parallel processor, providing a valuable tool for straight-line detection.

Thesis advisor, Douglas J. Fouts

The Final Proceedings for VECPAR 983rd International Meeting on Vector and Parallel Processing, 21 June 1998 -23 June 1998 This is an interdisciplinary conference. Topics include parallel and distributed computing, image processing and synthesis. mal-time and embedded systems.

The Final Proceedings for VECPAR, 4th International Meeting on Vector and Parallel Processing, 21-23 June 2000, an interdisciplinary conference covering topics in all areas of vector, parallel and distributed computing applied to a broad range of research disciplines with a focus on engineering. The principal topics include: Cellular Automata, Computational Fluid Dynamics, Crash and Structural Analysis, Data Warehousing and Data Mining, Distributed Computing and Operating Systems, Fault Tolerant Systems, Imaging and Graphics, Interconnection Networks, Languages and Tools, Numerical Methods, Parallel and Distributed Algorithms, Real-time and Embedded Systems, Reconfigurable Systems, Linear Algebra Algorithms and Software for Large Scientific Problems, Computer Organization, Image Analysis and Synthesis, and Nonlinear Problems.

We analyze the effects of hardware faults on the performance of computer-implemented signal detectors, as measured by the probability of detection and the probability of false alarm. We then use these results to design fault-tolerant detectors using hardware redundancy. Keywords include. Fault tolerance, Signal Processing, and Parallel Architecture.

Clustering data is one of the most common statistical and machine learning techniques for analyzing big data. Clustering can be particularly difficult when the data sets include categorical, missing, or noise variables. The tree clustering algorithm developed by Samuel Buttrey and Lyn Whitaker, as described in the December 2015 issue of The R Journal, seems to provide a solution to these problems, but it requires a large set of overhead computations. This issue is intensified when working with high-dimensional data because the extent of treeClust’s overhead computations are based on the dimensions of the data. High performance computing (HPC) and parallel processing present a solution to this overhead computation burden, but treeClust’s existing parallel processing method does not work on the Naval Postgraduate School’s HPC, the Hamming Supercomputer (HSC). Furthermore, correctly determining what HPC resources to use can be a difficult task. In this thesis, we present a new HSC-specific method for parallel processing data using the treeClust R package developed by Buttrey and Whitaker. Based on the results of our experiments, our method approximates the optimal resource HPC request, so that users realize the best run time when using treeClust on the HSC.

Clustering data is one of the most common statistical and machine learning techniques for analyzing big data. Clustering can be particularly difficult when the data sets include categorical, missing, or noise variables. The tree clustering algorithm developed by Samuel Buttrey and Lyn Whitaker, as described in the December 2015 issue of The R Journal, seems to provide a solution to these problems, but it requires a large set of overhead computations. This issue is intensified when working with high-dimensional data because the extent of treeClust’s overhead computations are based on the dimensions of the data. High performance computing (HPC) and parallel processing present a solution to this overhead computation burden, but treeClust’s existing parallel processing method does not work on the Naval Postgraduate School’s HPC, the Hamming Supercomputer (HSC). Furthermore, correctly determining what HPC resources to use can be a difficult task. In this thesis, we present a new HSC-specific method for parallel processing data using the treeClust R package developed by Buttrey and Whitaker. Based on the results of our experiments, our method approximates the optimal resource HPC request, so that users realize the best run time when using treeClust on the HSC.

A study was made of the emerging research area of algorithmic fault tolerance. This topic is attracting lots of attention due to its low implementation overhead and to its suitability for very large scale integration multiprocessor arrays. A fundamental contribution was made in the area by providing a novel linear algebraic model for the technique. In addition, a new fault tolerance scheme was developed for recursive least squares minimization. Other results include new ways for generating weights for the checksum method. (kr)

Includes bibliographical references (p. 391-393) and indexes

Includes bibliographical references (p. 391-393) and indexes

Reproducibility review completed, adding a registration for proper archival and snapshot.

In this paper, we study the communication complexity for the problem of computing a conjunctive query on a large database in a parallel setting with $p$ servers. In contrast to previous work, where upper and lower bounds on the communication were specified for particular structures of data (either data without skew, or data with specific types of skew), in this work we focus on worst-case analysis of the communication cost. The goal is to find worst-case optimal parallel algorithms, similar to the work of [18] for sequential algorithms. We first show that for a single round we can obtain an optimal worst-case algorithm. The optimal load for a conjunctive query $q$ when all relations have size equal to $M$ is $O(M/p^{1/\psi^*})$, where $\psi^*$ is a new query-related quantity called the edge quasi-packing number, which is different from both the edge packing number and edge cover number of the query hypergraph. For multiple rounds, we present algorithms that are optimal for several classes of queries. Finally, we show a surprising connection to the external memory model, which allows us to translate parallel algorithms to external memory algorithms. This technique allows us to recover (within a polylogarithmic factor) several recent results on the I/O complexity for computing join queries, and also obtain optimal algorithms for other classes of queries.

REAPER (DAW) instructional video: Parallel Processing