Good Speedup Research Articles

Scalable High-Quality Hypergraph Partitioning

Balanced hypergraph partitioning is an NP-hard problem with many applications, e.g., optimizing communication in distributed data placement problems. The goal is to place all nodes across k different blocks of bounded size, such that hyperedges span as few parts as possible. This problem is well-studied in sequential and distributed settings, but not in shared-memory. We close this gap by devising efficient and scalable shared-memory algorithms for all components employed in the best sequential solvers without compromises with regards to solution quality. This work presents the scalable and high-quality hypergraph partitioning framework Mt-KaHyPar . Its most important components are parallel improvement algorithms based on the FM algorithm and maximum flows, as well as a parallel clustering algorithm for coarsening – which are used in a multilevel scheme with log (n) levels. As additional components, we parallelize the n -level partitioning scheme, devise a deterministic version of our algorithm, and present optimizations for plain graphs. We evaluate our solver on more than 800 graphs and hypergraphs, and compare it with 25 different algorithms from the literature. Our fastest configuration outperforms almost all existing hypergraph partitioners with regards to both solution quality and running time. Our highest-quality configuration achieves the same solution quality as the best sequential partitioner KaHyPar , while being an order of magnitude faster with ten threads. Thus, two of our configurations occupy all fronts of the Pareto curve for hypergraph partitioning. Furthermore, our solvers exhibit good speedups, e.g., 29.6x in the geometric mean on 64 cores (deterministic), 22.3x (log (n) -level), and 25.9x ( n -level).

DKGR: A Distributed Framework for Geometric Knowledge Graph Embedding with Ray

In recent years, researchers were interested in learning knowledge graph (KG) embeddings because of their advantages in downstream tasks like link prediction, entity disambiguation, and entity Classification in KGs. The state-of-the-art KG embedding models are mainly implemented in a centralized environment, i.e., they are trained on a single machine. However, most of the real-world KGs are huge in size, and a single machine is not enough to store the whole KG and train embedding models in this context. There are a few frameworks for learning KG embeddings in distributed environments. In this paper, we propose a novel distributed framework, namely DKGR, for training geometric embedding models. Our framework uses the Ray platform as a distributed environment. We evaluate our framework for the link prediction task on real-world datasets. Our preliminary results show a good speed-up when training KG embedding models in our framework. They also show that DKGR is scalable to large-scale KGs without affecting the link prediction performance remarkably.

On the behavior of parallel island models

Parallel island models (PIMs) are used to enhance the performance of evolutionary algorithms (EAs). In PIMs, each island executes an EA for evolving its local population, and periodically individuals migrate to neighborhoods synchronously or asynchronously. Neighborhoods are organized through a topology of communication, and migration policies guide individual exchange. This work explores migration policies over different communication topologies in synchronous and asynchronous PIMs to improve the speed-up and accuracy of genetic algorithms (GAs). The aim is to explain such models’ adequacy from a general perspective, attempting to answer questions such as the best way to implement GAs in PIMs. To reach this goal, the quality of the solutions and the running time provided by proposed PIMs are evaluated over four NP-hard problems of different combinatorial nature: reversal and translocation evolutionary distance, task mapping and scheduling, and N-Queens. The experiments show that tuning the parameters of the breeding cycle and migration policies is vital to guarantee that all proposed PIMs reach good speed-ups and more accurate solutions than the sequential GA. In addition, experiments ratify that synchronous models provide the best solutions, while asynchronous models deliver the best speed-ups. Finally, the results show that no model provides either better speed-up or accuracy in general since, as for sequential models, the nature of the problem define which would be the best-adapted PIM.

Exploiting the Abstract Calculus Pattern for the Integration of Ordinary Differential Equations for Dynamics Systems: An Object-Oriented Programming Approach in Modern Fortran

This manuscript relates to the exploiting of the abstract calculus pattern (ACP) for the (numerical) solution of ordinary differential equation (ODEs) systems, which are ubiquitous mathematical formulations of many physical (dynamical) phenomena. We present FOODIE, a software suite aimed to numerically solve ODE problems by means of a clear, concise, and efficient abstract interface. The results presented prove manifold findings, in particular that our ACP approach enables ease of code development, clearness and robustness, maximization of code re-usability, and conciseness comparable with computer algebra system (CAS) programming (interpreted) but with the computational performance of compiled programming. The proposed programming model is also proven to be agnostic with respect to the parallel paradigm of the computational architecture: the results show that FOODIE applications have good speedup with both shared (OpenMP) and distributed (MPI, CAF) memory architectures. The present paper is the first announcement of the FOODIE project: the current implementation is extensively discussed, and its capabilities are proved by means of tests and examples.

Uncoupling system and environment simulation cells for fast-scaling modeling of complex continuum embeddings.

Continuum solvation models are becoming increasingly relevant in condensed matter simulations, allowing to characterize materials interfaces in the presence of wet electrified environments at a reduced computational cost with respect to all atomistic simulations. However, some challenges with the implementation of these models in plane-wave simulation packages still persists, especially when the goal is to simulate complex and heterogeneous environments. Among these challenges is the computational cost associated with large heterogeneous environments, which in plane-wave simulations has a direct effect on the basis-set size and, as a result, on the cost of the electronic structure calculation. Moreover, the use of periodic simulation cells is not well-suited for modeling systems embedded in semi-infinite media, which is often the case in continuum solvation models. To address these challenges, we present the implementation of a double-cell formalism, in which the simulation cell used for the continuum environment is uncoupled from the one used for the electronic-structure simulation of the quantum-mechanical system. This allows for a larger simulation cell to be used for the environment, without significantly increasing computational time. In this work, we show how the double-cell formalism can be used as an effective periodic boundary conditions correction scheme for nonperiodic and partially periodic systems. The accuracy of the double-cell formalism is tested using representative examples with different dimensionalities, both in vacuum and in a homogeneous continuum dielectric environment. Fast convergence and good speedups are observed for all the simulation setups, provided the quantum-mechanical simulation cell is chosen to completely fit the electronic density of the system.

Scalable and portable computational framework enabling online probabilistic remaining useful life (RUL) estimation

This work demonstrates a framework that enables online prognostics in adhesive joints by estimating the real-time probabilistic remaining useful life (RUL) using ANNs based hybrid physics models and vectorized sequential Monte Carlo (SMC) simulations. The framework is developed by integrating the physics-based damage degradation model and uncertainty quantification (UQ) techniques to estimate both probabilistic fatigue failure life and RUL. The fatigue damage growth (FDG) simulator, a hybrid surrogate model that simulates real-time fatigue degradation in adhesive joints, is used. In the initial set of results, the generalized framework is validated by estimating the probabilistic fatigue failure life using two UQ methods: Markov Chain Monte Carlo (MCMC) and SMC method. The computational results are successfully compared against experimental data. The conventional MCMC sampling methods are inherently serial, which limits the exploitation of the computational speed-up provided by the FDG simulator and hinders the real-time life predictions. The SMC method quantifies the uncertainties by parallelizing the sampling process, significantly reducing computational time and enabling real-time prediction. Next, the generalized framework is used to estimate probabilistic RUL from the fatigue crack propagation data. The parallel SMC method showed very good speedup compared to the MCMC method. To further enhance the computational speed-up with SMC method, vectorized FDG simulations are introduced into the framework and good scalability is achieved. Finally, the portability of the framework is demonstrated by deploying it on the portable Raspberry Pi cluster.

Odyssey: A Journey in the Land of Distributed Data Series Similarity Search

This paper presents Odyssey, a novel distributed data-series processing framework that efficiently addresses the critical challenges of exhibiting good speedup and ensuring high scalability in data series processing by taking advantage of the full computational capacity of modern distributed systems comprised of multi-core servers. Odyssey addresses a number of challenges in designing efficient and highly-scalable distributed data series index, including efficient scheduling, and load-balancing without paying the prohibitive cost of moving data around. It also supports a flexible partial replication scheme, which enables Odyssey to navigate through a fundamental trade-off between data scalability and good performance during query answering. Through a wide range of configurations and using several real and synthetic datasets, our experimental analysis demonstrates that Odyssey achieves its challenging goals.

Distributed random vector functional link network with subspace-based local connections

In order to solve the problem of poor generalization ability and high computational complexity of random vector functional link (RVFL) network when dealing with large-scale data classification, we design and implement a distributed RVFL network with subspace-based local connections in Spark framework (DRVFL-SLC). Firstly, in order to take advantage of the partition parallelism of resilient distributed dataset (RDD), the large-scale dataset stored in the Hadoop distributed file system HDFS is randomly divided into random sample partition (RSP) data blocks and each RSP data block corresponds to a partition of the RDD, where the RSP data block is a subset of data that maintains probability distribution consistency with the big data at a given significance level. After that, the mapPartitions transformation is invoked on the RDD containing multiple partitions in a distributed environment and this operation trains the corresponding optimal RVFL-SLC efficiently in parallel. Then, the collect execution operator is used to efficiently fuse the optimal RVFL-SLC corresponding to each partition of the RDD to obtain DRVFL-SLC for realizing the classification of big data. Finally, the feasibility and effectiveness of DRVFL-SLC are verified based on several large-scale data set with at least million records on a Spark cluster deployed with 6 computing nodes. The results show that DRVFL-SLC has a good speedup ratio, scalability and scale growth, and can achieve better generalization performance than RVFL-SLC trained on a single machine with full data.

Parallel Weighted Random Sampling

Data structures for efficient sampling from a set of weighted items are an important building block of many applications. However, few parallel solutions are known. We close many of these gaps. We give efficient, fast, and practicable parallel and distributed algorithms for building data structures that support sampling single items (alias tables, compressed data structures). This also yields a simplified and more space-efficient sequential algorithm for alias table construction. Our approaches to sampling k out of n items with/without replacement and to subset (Poisson) sampling are output-sensitive , i.e., the sampling algorithms use work linear in the number of different samples. This is also interesting in the sequential case. Weighted random permutation can be done by sorting appropriate random deviates. We show that this is possible with linear work. Finally, we give a communication-efficient, highly scalable approach to (weighted and unweighted) reservoir sampling. This algorithm is based on a fully distributed model of streaming algorithms that might be of independent interest. Experiments for alias tables and sampling with replacement show near linear speedups using up to 158 threads of shared-memory machines. An experimental evaluation of distributed weighted reservoir sampling on up to 5,120 cores also shows good speedups.

Intelligent Analysis of Logistics Information Based on Dynamic Network Data

In order to realize the intellectualization of logistics information analysis, this paper proposes an intelligent analysis method of logistics information based on dynamic network data cloud mining. This paper selects the data of a shipping logistics platform to realize the intelligent analysis experiment of logistics information based on cloud clustering mining. The purpose of the experiment is to find out the advantages of logistics information intelligent analysis based on cloud mining by comparing the performance differences between cloud clustering mining and traditional clustering mining in logistics information intelligent analysis. This paper builds an experimental environment based on Hadoop and MapReduce parallelization based on K-means algorithm. Taking the obtained logistics data as the analysis object, preprocess it and get the results based on cloud clustering mining. The experimental results show that the parallel mining analysis method is 179.2% slower than the traditional mining analysis method in dataset data1, 60.4% slower in dataset data2, and 2.8% faster in dataset data1. The intelligent analysis method of logistics information based on cloud clustering mining has good scalability and speedup ratio. Conclusion. Applying cloud mining to logistics information analysis and realizing the intelligent analysis of logistics information has great advantages, and can well meet the content and efficiency needs of logistics information analysis stakeholders.

Information-based massive data retrieval method based on distributed decision tree algorithm

Based on the distributed decision tree algorithm, this paper first proposes a method of vertically partitioning datasets and synchronously updating the hash table to establish an information-based mass data retrieval method in a heterogeneous distributed environment, as well as using interval segmentation and interval filtering technologies for improved algorithm of distributed decision tree. The distributed decision tree algorithm uses the attribute histogram data structure to merge the category list into each attribute list, reducing the amount of data that needs to reside in the memory. Second, we adopt the strategy of vertically dividing the dataset and synchronously updating the hash table, select the hash table entries that can be used to update according to the minimum Gini value, modify the corresponding entries and use the hash table to record and control each sub-site. In the case of node splitting, it has a high accuracy rate. In addition, for classification problems that meet monotonic constraints in a distributed environment, this paper will extend the idea of building a monotonic decision tree in a distributed environment, supplementing the distributed decision tree algorithm, adding a modification rule and modifying the generated nonmonotonic decision tree to monotonicity. In order to solve the high load problem of the privacy-protected data stream classification mining algorithm under a single node, a Storm platform for the parallel algorithm PPFDT_P based on the distributed decision tree algorithm is designed and implemented. At the same time, considering that the word vector model improves the deep representation of features and solves the problem of feature high-dimensional sparseness, and the iterative decision tree algorithm GBDT model is more suitable for non-high-dimensional dense features, the iterative decision tree algorithm will be integrated into the word vector model (GBDT) in the data retrieval application, using the distributed representation of words, namely word vectors, to classify short messages on the GBDT model. Experimental results show that the distributed decision tree algorithm has high efficiency, good speed-up and good scalability, so that there is no need to increase the number of datasets at each sub-site at any time. Only a small number of data items are inserted. By splitting some leaf nodes, a small amount is added by branching to achieve a monotonic decision tree. The proposed system achieves a massive data ratio of 54.1% while compared with other networks of massive data ratio.

Mining Order-preserving Submatrices under Data Uncertainty: A Possible-world Approach and Efficient Approximation Methods

Given a data matrix\( D \), a submatrix\( S \)of\( D \)is an order-preserving submatrix (OPSM) if there is a permutation of the columns of\( S \), under which the entry values of each row in\( S \)are strictly increasing. OPSM mining is widely used in real-life applications such as identifying coexpressed genes and finding customers with similar preference. However, noise is ubiquitous in real data matrices due to variable experimental conditions and measurement errors, which makes conventional OPSM mining algorithms inapplicable. No previous work on OPSM has ever considered uncertain value intervals using the well-established possible world semantics.We establish two different definitions of significant OPSMs based on thepossible world semantics: (1) expected support-based and (2) probabilistic frequentness-based. An optimized dynamic programming approach is proposed to compute the probability that a row supports a particular column permutation, with a closed-form formula derived to efficiently handle the special case of uniform value distribution and an accurate cubic spline approximation approach that works well with any uncertain value distributions. To efficiently check the probabilistic frequentness, several effective pruning rules are designed to efficiently prune insignificant OPSMs; two approximation techniques based on the Poisson and Gaussian distributions, respectively, are proposed for further speedup. These techniques are integrated into our two OPSM mining algorithms, based on prefix-projection and Apriori, respectively. We further parallelize our prefix-projection-based mining algorithm using PrefixFPM, a recently proposed framework for parallel frequent pattern mining, and we achieve a good speedup with the number of CPU cores. Extensive experiments on real microarray data demonstrate that the OPSMs found by our algorithms have a much higher quality than those found by existing approaches.

Parallel Domain Decomposition of a FEM-based Tool for Numerical Modelling Mineral Slurry-like Flows

The main parallelisation related features of a computational tool based on the finite element method (FEM) for the numerical modelling of mineral-slurry like flows are described in this work. In particular, both the domain decomposition method (DDM) and the processes communication strategy employed are discussed in detail. The DD algorithm is based on the iterative update of the boundary conditions imposed on the interfaces between subdomains, the so-called transmission conditions. Due to its versatility in several parallel architectures, the message-passing standard used here is the message passing interface (MPI) one. Since mineral-slurries rheology may change according to the prevailing local flow conditions, Newtonian and non-Newtonian viscous fluids are considered in this work. Indeed, both Newtonian and non-Newtonian laminar flows are numerically studied in two well-known canonical configurations usually found in mineral-slurry transport. The main results show that the parallel FEM based tool is capable of carrying out high-fidelity numerical simulations of mineral-slurry like flows. Finally, in all numerical simulations performed, relatively good speedups were obtained.

Parallel Cleaning Algorithm for Similar Duplicate Chinese Data Based on BERT

Data is an important source of knowledge discovery, but the existence of similar duplicate data not only increases the redundancy of the database but also affects the subsequent data mining work. Cleaning similar duplicate data is helpful to improve work efficiency. Based on the complexity of the Chinese language and the bottleneck of the single machine system to large-scale data computing performance, this paper proposes a Chinese data cleaning method that combines the BERT model and a k-means clustering algorithm and gives a parallel implementation scheme of the algorithm. In the process of text to vector, the position vector is introduced to obtain the context features of words, and the vector is dynamically adjusted according to the semantics so that the polysemous words can obtain different vector representations in different contexts. At the same time, the parallel implementation of the process is designed based on Hadoop. After that, k-means clustering algorithm is used to cluster similar duplicate data to achieve the purpose of cleaning. Experimental results on a variety of data sets show that the parallel cleaning algorithm proposed in this paper not only has good speedup and scalability but also improves the precision and recall of similar duplicate data cleaning, which will be of great significance for subsequent data mining.

Application of Gabor Image Recognition Technology in Intelligent Clothing Design

Aiming at the complex problem of image recognition feature extraction, this paper proposes an intelligent clothing design model based on parallel Gabor image feature extraction algorithm. Based on the intelligent parallel mode, the algorithm decomposes and merges the calculation process of the image Gabor transformation, decomposes the entire image Gabor feature extraction calculation process into a parallel part and a nonparallel part, and accelerates the parallel part by using multiple cores. The calculation results are then combined to achieve the purpose of multicore parallel acceleration of the entire calculation process. Secondly, based on the consideration of improving the real-time performance of the intelligent clothing design system, combined with the existing multicore environment, this paper uses the intelligent model to design and implement the image parallel Gabor feature extraction algorithm and uses image processing and analysis technology to analyze the visual elements of traditional clothing and identify and quantify to form a relatively complete clothing visual element evaluation system, which provides a basis for large-scale collection and automated evaluation of clothing visual effects, as well as clothing trend tracking and prediction. Experiments show that the algorithm can effectively shorten the calculation time of Gabor image feature extraction and can obtain a good speedup in a multicore environment. At the same time, it combines with a multiscale intelligent clothing classification algorithm, on the basis of the VS2008 platform, combined with OpenCV 2.0, designed and implemented an intelligent clothing design system, and conducted experiments and system tests. The experimental results show that the algorithm given in this paper can accurately segment fabric defects from the background, which proves that the detection algorithm has a good detection effect. Simulation results show that the algorithm proposed in this paper can more accurately identify the state of clothing features, and the real-time performance of intelligent clothing design in a multicore environment has been improved to a certain extent.

Speedup vs. quality: Asynchronous and cluster-based distributed adaptive genetic algorithms for ordered problems

While the main motivation for Parallel Genetic Algorithms (PGAs) has been to improve the scalability of Genetic Algorithms (GAs), techniques and strategies for maintaining population diversity is an equally active research topic. Island Model Genetic Algorithms (IMGAs) represent one of the most mature strategies for developing PGAs in an effective and scalable manner. However, identifying how much migration and which individuals should migrate are open research problems. Meanwhile, recent developments in Adaptive Genetic Algorithms (AGAs) have led to techniques for monitoring and maintaining population diversity in an online manner. The aim of the present work is to introduce adaptive techniques and mechanisms into PGAs in order to determine when, how much and which individuals are most suitable for migration. We present a number of adaptive PGAs that aim to maintain diversity and maximise coverage of the solution space by minimising the overlap between islands. PGAs presented in this work are empirically assessed for their abilities in scalability, ability to find good quality solutions and maintain population diversity in ordered problems. These metrics are compared to existing adaptive and parallel GAs selected from the literature for their performance. We estimated the overhead costs of monitoring diversity and communication would result in a trade off between scalability and search capabilities. Our results suggest that an asynchronous adaptive PGA has the greatest speedup potential. However, while localising adaptive populations by k-means clustering is less scalable, results indicate that the method is more effective at directing the search in order to reduce the likelihood of islands searching in the same areas of the solution space. For this reason, an adaptive PGA with clustering-based migration demonstrates greater potential in solution quality while maintaining good speedup performance.

A parallel multilevel domain decomposition method for source identification problems governed by elliptic equations

In this paper we develop a parallel multilevel domain decomposition method for large-scale source identification problems governed by elliptic equations. A popular approach is to formulate the inverse problem as a PDE-constrained optimization problem. The minima satisfies a Karush–Kuhn–Tucker (KKT) system consisting of the state, adjoint and source equations which is rather difficult to solve on parallel computers. We propose and study a parallel method that decomposes the optimization problem on the global domain into subproblems on overlapping subdomains, each subdomain is further decomposed to form an additive Schwarz preconditioner for solving these smaller subproblems simultaneously with a preconditioned Krylov subspace method. For each subproblem, the overlapping part of the solution is discarded and the remaining non-overlapping part of the solution is put together to obtain an approximated global solution to the inverse problem. Since all the subproblems are solved independently, the multilevel domain decomposition method has the advantage of higher degree of parallelism. Numerical experiments show that the algorithm is accurate in terms of the reconstruction error and has reasonably good speedup in terms of the computing time. The efficiency and robustness of the proposed approach on a parallel computer with more than 1, 000 processors are reported.

An efficient automated incremental density-based algorithm for clustering and classification

Data clustering divides the datasets into different groups. Incremental Density-Based Spatial Clustering of Applications with Noise (DBSCAN) is a famous density-based clustering technique able to find the clusters of variable sizes and shapes. The quality of incremental DBSCAN results has been influenced by two input parameters: MinPts (Minimum Points) and Eps (Epsilon). Therefore, the parameter setting is one of the major problems of incremental DBSCAN. In the present article, an improved incremental DBSCAN accorded to Non-dominated Sorting Genetic Algorithm II (NSGA-II) has been presented to address the issue. The proposed algorithm adjusts the two parameters (MinPts and Eps) of the incremental DBSCAN via the iteration and the fitness functions to enhance the clustering precision. Moreover, our proposed method introduces suitable fitness functions for both labeled and unlabeled datasets. We have also improved the efficiency of the proposed hybrid algorithm by parallelization of the optimization process. The evaluation of the introduced method has been done through some textual and numerical datasets with different shapes, sizes, and dimensions. According to the experimental results, the proposed algorithm provides better results than Multi-Objective Particle Swarm Optimization (MOPSO) based incremental DBSCAN and a few well-known techniques, particularly regarding the shape and balanced datasets. Also, good speed-up can be reached with a parallel model compared with the serial version of the algorithm.

Serial and parallel memetic algorithms for the bounded diameter minimum spanning tree problem

AbstractGiven a connected, weighted, undirected graph G = (V, E) and an integer D ≥ 2, the bounded diameter minimum spanning tree (BDMST) problem seeks a spanning tree of minimum cost, whose diameter is no greater than D. The problem is known to be NP‐hard, and finds application in various domains such as information retrieval, wireless sensor networks and distributed mutual exclusion. This article presents a two‐phase memetic algorithm for the BDMST problem that combines a specialized recombination operator (proposed in this work) with good heuristics in order to more effectively direct the exploration of the search space into regions containing better solutions. A parallel meta‐heuristic is also proposed and shown to obtain very good speedups – super‐linear, in several cases – vis‐à‐vis the serial memetic algorithm. To the best of the authors' knowledge, this is the first parallel meta‐heuristic proposed for the BDMST problem, and potentially paves the way for handling much larger problems than reported in the literature. Some observations and theorems are presented in order to provide the underlying framework for the proposed algorithms. Further, the proposed memetic algorithm and parallel meta‐heuristic are shown, in the course of several computational experiments, to in general obtain superior solution quality with much lesser computational effort in comparison to the best known meta‐heuristics in the literature over a wide range of benchmark instances.

Fast t-SNE algorithm with forest of balanced LSH trees and hybrid computation of repulsive forces

An acceleration of the well-known t-Stochastic Neighbor Embedding (t-SNE) (Hinton and Roweis, 2003; Maaten and Hinton, 2008) algorithm, probably the best (nonlinear) dimensionality reduction and visualization method, is proposed in this article.By using a specially-tuned forest of balanced trees constructed via locality sensitive hashing is improved significantly upon the results presented in Maaten (2014), achieving a complexity significantly closer to true O(nlogn), and vastly improving behavior for huge numbers of instances and attributes. Such acceleration removes the necessity to use PCA to reduce dimensionality before the start of t-SNE.Additionally, a fast hybrid method for repulsive forces computation (a part of the t-SNE algorithm), which is currently the fastest method known, is proposed.A parallelized version of our algorithm, characterized by a very good speedup factor, is proposed.