Density functional theory is a method for calculation of electronic structure of physical systems with a large number of atoms. In an effort to produce parallel implementation capable of solving systems with tens of thousands of atoms, in the first steps, we use Gaussian orbitals that allow relevant integrals to be calculated analytically. The main goal of this paper is the serial implementation of these integral calculations using C programming language. The analytical solutions consist of a very complex nested summations whose efficient implementation is the main contribution of this paper. We describe challenges encountered during implementation and their solutions. Results show that the given C implementation is at least three orders of magnitude faster than numerical solutions in the best available computational software programs.
This paper introduces our work on developing Krylov subspace and AMG solvers on NVIDIA GPUs. As SpMV is a crucial part for these iterative methods, SpMV algorithms for a single GPU and multiple GPUs are implemented. A HEC matrix format and a communication mechanism are established. Also, a set of specific algorithms for solving preconditioned systems in parallel environments are designed, including ILU(k), RAS and parallel triangular solvers. Based on these work, several Krylov solvers and AMG solvers are developed. According to numerical experiments, favorable acceleration performance is obtained from our Krylov solver and AMG solver under various parameter conditions.
Based on the characteristics of a huge number of data bits and reconfigurable optical processors of a ternary optical computer, a high-efficient division algorithm and realization are presented in this paper. The division algorithm adopts the SRT division to simplify operation steps and uses a two-step parallel M+B adder to speed up calculations. A number of rigorous tests for this algorithm are carried out through software simulation, which proves that the accuracy and performance of division in a ternary optical computer is superior to that of an electronic computer.
In the paper a new dispersion-relation-preserving upwind combined compact difference scheme (DRP-UCCD) to solve a time-dependent convection diffusion equations with variable coefficients is proposed. The developed scheme is constructed by making use of the high-order upwind combined compact difference operators, which can preserve the dispersion relation and enhance the convective stability. The scheme is proved to have the unconditional stability and the error accuracy is six order on space and two order on time. Numerical experiments confirm its high efficiency.
As one of the important techniques of data analysis, a clustering algorithm is widely used in data mining, image recognition, information extraction, pattern recognition and other fields. In the era of big data, with the rapid development of web applications, much of the data to be processed is characterized by massive and dynamic growth. Under this background, how to cluster the incremental data becomes a challenging problem that clustering algorithms face. In this paper, we proposed a limited incremental clustering algorithm with respect to cluster stability. Based on an assumption that the categories of data are limited, we take advantage of an existing clustering structure and deal with incremental data steadily with respect to cluster stability. Cluster reconstruction will be triggered when the stability does no longer hold or a buffer pool for undetermined data is full. At the end of the paper, we implement the limited incremental clustering algorithm with K-means. Meanwhile, we use an average density of clusters and the global stability to choose a proper value of K.
Transductive Support Vector Machine (TSVM) is a famous model in solving classification problems in semi-supervised learning (SSL). However, traditional methods of TSVM are always time-consuming, which becomes a bottleneck when there is large scale data. In order to improve the efficiency of a training process for TSVM, we propose a fast training method for classification problems in semi-supervised learning. Different from the traditional algorithms, just few labeled instances are regarded as the initial training dataset at the beginning of our method. In each following iteration, only reliable unlabeled instances are selected and then labels are set by using all instances which are the labels set by the previous iteration. The algorithm finishes until no reliable unlabeled instances can be selected. This core idea makes our proposed method consume less running time than traditional TSVM implementation on SVMlight. In addition, for different degrees of reliable unlabeled instances, different label setting strategies will be applied to keep high accuracy in label setting. Simulations on both artificial data and real-world data have proven that our proposed fast training method can greatly reduce the running time compared to traditional TSVM implementation on SVMlight. Meanwhile, it can reach the same accuracy level in classification.
A mobile sink has been exploited for data gathering in Wireless Sensor Networks (WSN), which can reduce and balance energy consumption among sensor nodes. A mobile sink assisted transmission algorithm has become one of the research hotspots for WSN. How to energy-efficiently collect and transmit data by using multiple mobile sinks is one of the hot research topics. In this paper, we propose a mobile smart UE-gateway assisted transmission algorithm for WSN, which considers several parameters to choose an optimized mobile smart UE-gateway as an access point. Then we analyze the WSN system energy cost and transmission delay. The performance evaluation results show that system energy costs, WSNs lifetime and transmission delay can be improved largely by using the optimization of a mobile smart UE-gateway.
This paper addresses the anonymous communications technology. Multiple hops, self-organization, no fixed facilities and limited computing resources of wireless ad hoc networks (MANET) make it hard to apply high complexity of a traditional network security algorithm. Security mechanisms can better solve the security issues of anonymous communication. This paper mainly studies the existing anonymous technology of traditional networks and MANET, and summarizes its advantages and disadvantages. With this article, readers can have a more thorough understanding of anonymous communication security and research trends in this area.
The lattice Boltzmann Method (LBM), different from classical numerical methods of continuum mechanics, is derived from molecular dynamics. The LBM has the following main advantages: including a simple algorithm, the direct solver for pressure, easy treatment of complicated boundary conditions and particularly parallel suitability. The most common models include the Single-Relaxation-Time (SRT) and Multiple-Relaxation-Time (MRT) collision models. In a conventional parallel computing model of LBM, communication and computing are performed individually. When the communication is performed, the computing is waiting in MPI processes. This will waste some waiting time. Therefore, the communication and computing overlapping parallel model was proposed. By the architecture of "Ziqiang 4000" supercomputer at Shanghai University, the hybrid MPI and OpenMP parallel model is proposed. The numerical results show that the presented model has better computational efficiency.
In this paper, we study the parallelization of a Cartesian grid based treecode algorithm in evaluating electrostatic potentials in a charged particle system. The treecode algorithm uses a far-field Taylor expansion to compute 𝒪(N logN) particle-cluster interactions to replace the 𝒪(N^2) particle-particle interactions. The treecode algorithm is implemented with MPI based parallelization. We design schemes to optimize the implementation adaptive to the particle location. The numerical results show high parallel efficiency. These optimized schemes are further extended to accelerate GMRES iteration in solving boundary integral Poisson–Boltzmann equation in which the discretized linear algebraic system resembles the interactions of the charged system.
Time for swimming at young ages may be a good indicator for swimmers’ future performance. Through analyzing a large data set on swimming time, we use the machine learning algorithms to explore swimmers’ performance on four different strokes in a 100 m long course for both males and females. For each stroke, we divide swimmers’ performance into four levels according to their time at the ages of 12–13, and predict their performance levels at the age of 18 using two well-known machine learning methods with optimal parameters. Based on the existing data, we predict the probability from each level at a young age to the top 25
The continuous development of the information technology leads to the explosive growth of many information domains. Obtaining the required information from a large-scale text in a quick and accurate way has become a great challenge. Keyword extraction is a kind of effective method to solve these problems. It is one of the core technologies in the research area of text mining, and plays a very important role. Currently, the keywords of most text information have not been provided. Some keywords of a text are not contained in the text content. There is not any elegant solution, offered by the existing algorithms, for this problem yet. To solve it, this paper proposes a keyword extraction method based on word vectors. The concept of a text turns into computer understandable space by training word vectors using a word2vec algorithm. This method trains all the words and keywords which appear in the text into vector sets through the word2vec training method, and then the words in the test text will be replaced by word term vectors. The Euclidean distances between every candidate words and every text words are calculated to find out the top-N-closest keywords as the automatic text extraction keywords. The experiment uses computer field papers as a training text. The results show that the method can improve the accuracy of the phrase keyword extraction and find the keywords not appearing in the text.
A ternary optical computer (TOC) has three prominent features, numerous data bits, reconstruction of each bit's function and allocation bit by bit. The numerous data bits determine that a TOC is particularly suitable to processing large amounts of data, especially in high-performance computing where data is more and more. In order to organize large amounts of data which is entered into a TOC, a new kind of file is defined, which includes all input operands and calculating rules. The file is called SZG file because it has a suffix. SZG. The paper introduces the demands, significance, construction principles and usage method of a SZG file. Finally, an example of using a SZG file in high-performance computing is given.
This paper presents our work on simulation of large-scale reservoir models on IBM Blue Gene/Q and studying the scalability of our parallel reservoir simulators. An in-house black oil simulator has been implemented. It uses MPI for communication and is capable of simulating reservoir models with hundreds of millions of grid cells. Benchmarks show that our parallel simulators are thousands of times faster than sequential simulators that are designed for workstations and personal computers, and these simulators have excellent scalability.
According to the conventional Maxwell distribution function, a new equilibrium distribution function based on a discrete velocity model (D2Q13) is proposed. A parallel lattice Boltzmann algorithm based on this new function is used for simulating the lid-driven cavity flow. The experimental results validate the correctness of the new equilibrium distribution function.
Absolute value equations $$ Ax - |x| = b $$ are non-differentiable hard problems. Many linear and quadratic programming problems can ultimately be converted into absolute value equation problems so research on solving an absolute value problem has important practical and theoretical significance. An improved adaptive differential evolution algorithm was proposed to solve the absolute value equations in this paper. The algorithm combined global search ability and local search ability, using an adaptive quadratic mutation operation and crossover operation. Numerical results show that the improved algorithm can quickly find the solutions of these equations.
Since the classic determinant computation method Cholesky decomposition may devastate sparsity of matrices and cost cubic steps, it is impractical to apply this method to large-scale symmetric positive-definite matrices due to limitation of storage and efficiency. Therefore, a randomized algorithm is proposed to calculate log-determinants of symmetric positive-definite matrices via stochastic trace approximations, implemented by weighted L-2 orthogonal polynomial expansions with efficient recursion formulas and matrix-vector multiplications based on the matrix eigenvalue distribution. As Chebyshev expansions have been applied to this problem before, our main contribution is proposing the strategies of weighted function selection based on prior eigenvalue distribution, which generalizes approximating polynomials for this problem and may accelerate computation.
Traditionally, solving the parameter identification inverse problems of partial differential equations encountered many difficulties and insufficiency. In this paper, we propose an improved GEP (Gene Expression Programming) to identify the parameters in the reverse problems of partial differential equations based on the self-adaption, self-organization and self-learning characters of GEP. This algorithm simulates a parametric function itself of a partial differential equation directly through the observed values by fully taking into account inverse results caused by noises of a measured value. Modeling is unnecessary to add regularization in the modeling process aiming at special problems again. The experiment results show that the algorithm has good noise-immunity. In case there is no noise or noise is very low, the identified parametric function is almost the same as the original accurate value; when noise is very high, good results can still be obtained, which successfully realizes automation of the parameter modeling process for partial differential equations.
The Solar Broadband Radio Spectrometer (SBRS) monitors the solar radio busts all day long and produces solar radio astronomical big data foranalysis every day, which usually have been accumulated in mass images for scientific study over decades. In the observed mass data, burst events are rare and always along with interference, so it seems impossible to identify whether the mass data contain bursts or not and figure out which type of burst it is by manual operation timely. Therefore, we take advantage of high performance computing and machine learning techniques to classify the huge volume astronomical imaging data automatically. The professional line of multiple NVIDIA GPUs has been exploited to deliver 78x faster parallel processing power for high performance computing of the astronomical big data, and neural networks have been utilized to learn the representations of the solar radio spectra. Experimental results have demonstrated that the employed network can effectively classify a solar radio image into the labeled categories. Moreover, the processing time is dramatically reduced by exploring GPU parallel computing environment.
A ternary optical computer (TOC) has many advantages compared with its electronic counterpart such as huge data bits, excellent parallelism, and reconfigurable data bit. However, traditionally, to use a TOC, the user has to get to know some details about it. This does not benefit to exert the advantages of TOCs. One of the most attractive solutions is to develop a kind of platform for routine development by shielding the underlying hardware and let the user uses a TOC just like using electronic computers without needing to understand the details of an optical system. As a preliminary attempt, we put forward the idea of how to design an application routine platform and present the implementation process of the platform. Meanwhile, we design a basic version of the application routine platform and execute a MSD parallel addition routine on it. Experiment results show that the routine platform is correct and feasible. It can mask the details about a TOC and the internal operational logic of the platform can quite exert the advantages of the TOC.
