Network Of Workstations Research Articles

Parallel Fast Isogeometric Solvers for Explicit Dynamics

This paper presents a parallel implementation of the fast isogeometric solvers for explicit dynamics for solving non-stationary time-dependent problems. The algorithm is described in pseudo-code. We present theoretical estimates of the computational and communication complexities for a single time step of the parallel algorithm. The computational complexity is O(p^6 N/c t_comp) and communication complexity is O(N/(c^(2/3)t_comm) where p denotes the polynomial order of B-spline basis with Cp-1 global continuity, N denotes the number of elements and c is number of processors forming a cube, t_comp refers to the execution time of a single operation, and t_comm refers to the time of sending a single datum. We compare theoretical estimates with numerical experiments performed on the LONESTAR Linux cluster from Texas Advanced Computing Center, using 1 000 processors. We apply the method to solve nonlinear flows in highly heterogeneous porous media.

Transforming a Local Medical Image Analysis for Running on a Hadoop Cluster

There is a progressive digitization in many medical fields, such as digital microscopy, which leads to an increase in data volume and processing demands for the underlying computing infrastructure. This paper explores scaling behaviours of a Ki–67 analysis application, which processes medical image tiles, originating from a WSI (Whole slide Image) file format. Furthermore, it describes how the software is transformed from a Windows PC to a distributed Linux cluster environment. A test for platform independence revealed a non-deterministic behaviour of the application, which has been fixed successfully. The speedup of the application is determined. The slope of the increase is quite close to 1, i.e. there is almost no loss due to a parallelization overhead. Beyond the cluster’s hardware limit (72 cores, 144 threads, 216 GB RAM) the speedup saturates to a value around 64. This is a strong improvement of the original software, whose speedup is limited to two.

An UHD video handling system using a scalable server over an IP network

Ultra-high definition (UHD) videos, such as 4K and 8K, have four times higher resolution than current HDTV videos. Broadcasting stations want to deliver these videos with as little loss as possible. However, uncompressed UHD video handling systems, such as storage and transmission systems, on IP networks are not available. We researched and developed a scalable high-speed video server system to store and deliver HD or 4K videos without compression by adapting a PC cluster technology. To handle videos with higher resolution than 4K, we also developed a method that synchronises with multiple the servers and 4K video transmission systems. We achieved 54 Gbit/s delivery performance with single cluster server comprising 24 PCs. We also implemented the synchronisation method on the server and QG-70 video transmission system. As a result, we were able to demonstrate the world's first bi-directional, real-time, uncompressed 8K-video transmission system and the 8K video-on-demand system.

An UHD video handling system using a scalable server over an IP network

Ultra-high definition (UHD) videos, such as 4K and 8K, have four times higher resolution than current HDTV videos. Broadcasting stations want to deliver these videos with as little loss as possible. However, uncompressed UHD video handling systems, such as storage and transmission systems, on IP networks are not available. We researched and developed a scalable high-speed video server system to store and deliver HD or 4K videos without compression by adapting a PC cluster technology. To handle videos with higher resolution than 4K, we also developed a method that synchronises with multiple the servers and 4K video transmission systems. We achieved 54 Gbit/s delivery performance with single cluster server comprising 24 PCs. We also implemented the synchronisation method on the server and QG-70 video transmission system. As a result, we were able to demonstrate the world's first bi-directional, real-time, uncompressed 8K-video transmission system and the 8K video-on-demand system.

Optimized R functions for analysis of ecological community data using the R virtual laboratory (RvLab).

BackgroundParallel data manipulation using R has previously been addressed by members of the R community, however most of these studies produce ad hoc solutions that are not readily available to the average R user. Our targeted users, ranging from the expert ecologist/microbiologists to computational biologists, often experience difficulties in finding optimal ways to exploit the full capacity of their computational resources. In addition, improving performance of commonly used R scripts becomes increasingly difficult especially with large datasets. Furthermore, the implementations described here can be of significant interest to expert bioinformaticians or R developers. Therefore, our goals can be summarized as: (i) description of a complete methodology for the analysis of large datasets by combining capabilities of diverse R packages, (ii) presentation of their application through a virtual R laboratory (RvLab) that makes execution of complex functions and visualization of results easy and readily available to the end-user.New informationIn this paper, the novelty stems from implementations of parallel methodologies which rely on the processing of data on different levels of abstraction and the availability of these processes through an integrated portal. Parallel implementation R packages, such as the pbdMPI (Programming with Big Data – Interface to MPI) package, are used to implement Single Program Multiple Data (SPMD) parallelization on primitive mathematical operations, allowing for interplay with functions of the vegan package. The dplyr and RPostgreSQL R packages are further integrated offering connections to dataframe like objects (databases) as secondary storage solutions whenever memory demands exceed available RAM resources.The RvLab is running on a PC cluster, using version 3.1.2 (2014-10-31) on a x86_64-pc-linux-gnu (64-bit) platform, and offers an intuitive virtual environmet interface enabling users to perform analysis of ecological and microbial communities based on optimized vegan functions.A beta version of the RvLab is available after registration at: https://portal.lifewatchgreece.eu/

A MPI + OpenMP + CUDA Hybrid Parallel Scheme for MT Occam Inversion

To improve the performance of the Magnetotelluric Occam inversion, by in-depth analysis of the sequential algorithm, we develop a multi-level hybrid parallel computing scheme for MT Occam inversion based on MPI+OpenMP+CUDA and implement it on a small heterogeneous cluster. We implement the parallel algorithm for solving linear equations with Gauss elimination, jacobian matrix, cross-product matrix calculations and Cholesky decomposition. Through reasonable decomposition, combination and mapping of computing tasks, the scheme reduces the data traffic and realizes the purpose of load balancing. By changing the matrix storage order， the memory access speed is significantly increased. The scheme is tested with multiple synthesis data from 2-D theoretical models and the execution efficiency of sequential code and parallel code on a 4 nodes PC cluster is comparatively analyzed. The test results show that the realization of this hybrid parallel algorithm is feasible and efficient. Compared with the sequential code and pure message passing algorithm, the inversion speed is obviously increased.

Fast Optimal Design of Micromagnetic Devices Using FastMag and Distributed Evolutionary Algorithm

A new software framework is presented for fast optimal designs of micromagnetic devices. The developed framework adopts a fast micromagnetic (FastMag) simulator as a micromagnetic solver, and utilizes distributed evolutionary algorithm in Python (DEAP). The framework makes use of a simple Linux utility for resource management (SLURM) in order to run and manage multiple micromagnetic simulations concurrently on a Linux cluster. Various design variables can be designated including material and geometry parameters. The whole design cycle is automated including the mesh generation, and the status of the evolution and multiple simulations can be monitored and managed easily. By combining high computing speed of FastMag and distributed optimization capability of DEAP and SLURM, the optimal designs of micromagnetic devices can be performed rapidly and efficiently. As numerical examples, the optimal designs of a spin-torque oscillator and a magnetic random access memory device are performed.

Cljam: a library for handling DNA sequence alignment/map (SAM) with parallel processing.

BackgroundNext-generation sequencing can determine DNA bases and the results of sequence alignments are generally stored in files in the Sequence Alignment/Map (SAM) format and the compressed binary version (BAM) of it. SAMtools is a typical tool for dealing with files in the SAM/BAM format. SAMtools has various functions, including detection of variants, visualization of alignments, indexing, extraction of parts of the data and loci, and conversion of file formats. It is written in C and can execute fast. However, SAMtools requires an additional implementation to be used in parallel with, for example, OpenMP (Open Multi-Processing) libraries. For the accumulation of next-generation sequencing data, a simple parallelization program, which can support cloud and PC cluster environments, is required.ResultsWe have developed cljam using the Clojure programming language, which simplifies parallel programming, to handle SAM/BAM data. Cljam can run in a Java runtime environment (e.g., Windows, Linux, Mac OS X) with Clojure.ConclusionsCljam can process and analyze SAM/BAM files in parallel and at high speed. The execution time with cljam is almost the same as with SAMtools. The cljam code is written in Clojure and has fewer lines than other similar tools.

A Unified and Comprehensible View of Parametric and Kernel Methods for Genomic Prediction with Application to Rice.

One objective of this study was to provide readers with a clear and unified understanding of parametric statistical and kernel methods, used for genomic prediction, and to compare some of these in the context of rice breeding for quantitative traits. Furthermore, another objective was to provide a simple and user-friendly R package, named KRMM, which allows users to perform RKHS regression with several kernels. After introducing the concept of regularized empirical risk minimization, the connections between well-known parametric and kernel methods such as Ridge regression [i.e., genomic best linear unbiased predictor (GBLUP)] and reproducing kernel Hilbert space (RKHS) regression were reviewed. Ridge regression was then reformulated so as to show and emphasize the advantage of the kernel “trick” concept, exploited by kernel methods in the context of epistatic genetic architectures, over parametric frameworks used by conventional methods. Some parametric and kernel methods; least absolute shrinkage and selection operator (LASSO), GBLUP, support vector machine regression (SVR) and RKHS regression were thereupon compared for their genomic predictive ability in the context of rice breeding using three real data sets. Among the compared methods, RKHS regression and SVR were often the most accurate methods for prediction followed by GBLUP and LASSO. An R function which allows users to perform RR-BLUP of marker effects, GBLUP and RKHS regression, with a Gaussian, Laplacian, polynomial or ANOVA kernel, in a reasonable computation time has been developed. Moreover, a modified version of this function, which allows users to tune kernels for RKHS regression, has also been developed and parallelized for HPC Linux clusters. The corresponding KRMM package and all scripts have been made publicly available.

Performance analysis of a parallel algorithm for restoring large-scale CT images

In multiple areas of image processing, such as Computed Tomography, in which data acquisition is based on counting particles that hit a detector surface, Poisson noise occurs. Using variance-stabilizing transformations, the Poisson noise can be approximated by a Gaussian one, for which classical denoising filters can be used. This paper presents an experimental performance study of a parallel implementation of the Poissonian image restoration algorithm, introduced in Harizanov et al. (2013). Hybrid parallelization based on MPI and OpenMP standards is investigated. The convergence rate of the algorithm heavily depends on both the image size and the choice of input parameters (ρ,σ), thus maximizing its parallel efficiency is vital for real-life applications. The implementation is tested for high-resolution radiographic images, on Linux clusters with Intel processors and on an IBM supercomputer.

Optimization of Truss Structure Using Genetic Algorithm Performed on GPU

Modern Graphics Processing Units (GPU), offer a tremendous computing power, that is frequently an order of magnitude larger than even the most modern multi-core CPUs, making them an attractive platform for high performance computing due to their relative cheapness compared with conventional PC clusters. General purpose computing on GPUs (GPGPU) is becoming popular in High Performance Computing (HPC) because of its high peak performance. In this paper, a typical two-dimensional truss structure optimization problem is solved using Binary Genetic Algorithm (BGA) on both CPU and GPU. The kernel inside the GPU code computes the nodal displacements and elemental stresses by Finite Element Analysis (FEA) to evaluate the objective function and the constraints while making use of the Single Instruction Multiple Data (SIMD) structure of GPU to attain parallelization. The results are assessed for different values of parameters, such as complexity of the problem, number of elements, population size, maximum allowable generations and number of threads etc. to demonstrate how the value of speedup varies with these parameters and to provide a basic guideline for choosing the parameters for a different problem. The results clearly establish that calculations are performed considerably faster through the GPU than through the CPU in general.

BluePyOpt: Leveraging Open Source Software and Cloud Infrastructure to Optimise Model Parameters in Neuroscience.

At many scales in neuroscience, appropriate mathematical models take the form of complex dynamical systems. Parameterizing such models to conform to the multitude of available experimental constraints is a global non-linear optimisation problem with a complex fitness landscape, requiring numerical techniques to find suitable approximate solutions. Stochastic optimisation approaches, such as evolutionary algorithms, have been shown to be effective, but often the setting up of such optimisations and the choice of a specific search algorithm and its parameters is non-trivial, requiring domain-specific expertise. Here we describe BluePyOpt, a Python package targeted at the broad neuroscience community to simplify this task. BluePyOpt is an extensible framework for data-driven model parameter optimisation that wraps and standardizes several existing open-source tools. It simplifies the task of creating and sharing these optimisations, and the associated techniques and knowledge. This is achieved by abstracting the optimisation and evaluation tasks into various reusable and flexible discrete elements according to established best-practices. Further, BluePyOpt provides methods for setting up both small- and large-scale optimisations on a variety of platforms, ranging from laptops to Linux clusters and cloud-based compute infrastructures. The versatility of the BluePyOpt framework is demonstrated by working through three representative neuroscience specific use cases.

Robust and portable capacity computing method for many finite element analyses of a high-fidelity crustal structure model aimed for coseismic slip estimation

Computation of many Green's functions (GFs) in finite element (FE) analyses of crustal deformation is an essential technique in inverse analyses of coseismic slip estimations. In particular, analysis based on a high-resolution FE model (high-fidelity model) is expected to contribute to the construction of a community standard FE model and benchmark solution. Here, we propose a naive but robust and portable capacity computing method to compute many GFs using a high-fidelity model, assuming that various types of PC clusters are used. The method is based on the master–worker model, implemented using the Message Passing Interface (MPI), to perform robust and efficient input/output operations. The method was applied to numerical experiments of coseismic slip estimation in the Tohoku region of Japan; comparison of the estimated results with those generated using lower-fidelity models revealed the benefits of using a high-fidelity FE model in coseismic slip distribution estimation. Additionally, the proposed method computes several hundred GFs more robustly and efficiently than methods without the master–worker model and MPI.

Modelling of Complexities in PA of Laplace Equation

Dominant trends in parallel and distributed computing (parallel computers, parallel algorithms) aim to connect a number of powerful workstations (WS) based on single personal computers (PC) or symmetrical multiprocessor system (SMP) to solve complex problems in a parallel way. Typical representatives of these trends are network of workstation (NOW) and Grid as a high integrated network of NOW modules. In relation to it there is a bad need of unified modelling in traditionally evolved parallel computing and distributed computing as two separate research disciplines. Likewise current trends in high performance computing (HPC) aim to use NOW modules based on SMP parallel computers as cheaper alternatives to the traditionally used massive parallel multiprocessors or supercomputers and to profit from unifying modelling in parallel and distributed computing too. To exploit the parallel processing capability of such cluster, any application program must be paralleled. A way how to do it (decomposition model) belongs to a most important step in developing process of effective parallel algorithm. This article at first discusses the development of effective (optimised) iterative parallel algorithms (IPA) to solve partial differential equations (PDE). By carefully choosing an illustrative typical problem of PDE (Laplace equation) the proposed research has demonstrated the most important complexities in the process of complete modelling of substantial complexities and their optimisation in order to develop an effective PA. Such approach could be used as illustrative example for a complete modelling and effective solution of even other more complex PDE´s in a parallel way.

Modelling of Communication Complexity in Computers

We have been oft used to ignore the influence of communication in sequential computers or in better cases we have been supposed that this influence should be included in computation complexity representing performed I/O instructions. Multiple using of high performed CPU (central processor unit) within a single computer (SMP parallel computer) but also multiple using of high performed computers (workstations) as the building computing node of dominant parallel computers nowadays (network of workstations - NOW) or integrated network of NOW modules - Grid).The extended communication steps in parallel computers represent inter process communication (IPC) of decomposed parts of given parallel algorithm (PA) named as parallel processes. Then we can simply say that any PA consists of parallel processes with their sequential character and IPC among parallel processes. In relation to increasing role of communications in parallel computers this paper is devoted to modelling of communication complexity (number of communication steps) in the same way as we use to analyse computation complexity so in sequential algorithms. Based on this extended approach we are able better to optimise performance of all existed parallel computers. Another important problem is to analyse critical parts of communication influence to parallel computer performance and that at first in minimisation of communications at developing stages of any PA and secondly at its whole performance optimisation. Finally we have illustrated from the point of user the important role of communication in parallel computers on the chosen examples.

Optimisation of Communication Complexity in Parallel Computing

Parallel principles have been for a long time the most effective way how to increase performance in parallel computing (parallel computers, parallel algorithms). Dominant parallel computers are based on network of workstations (NOW) and on high integrated network of NOW modules (Grid). Effective using of such parallel computers assumes minimisation of at least substantial overheads representing with overhead function h(s, p). To the most important overhead latency belongs communication latency representing by inter process communication (IPC) of decomposed parallel processes. In this sense the paper is devoted to the modelling of communication complexity in unified parallel and distributed computing. This overhead function, representing by communication overhead, is then an integral part of the complex parallel execution time. Based on this we are able better to optimise PA during the developing process of PA and to come to final effective (optimised) PA. The article analysis such parts in PA developing process which are critical to communication complexity and that first the critical role of problem parallelisation (decomposition strategy) and second at own performance PA optimisation (shared memory, distributed memory, hybrid). Based on analysed optimised examples of PA the paper illustrates these critical parts from the point of user. In similar way the article point to concept of complex isoefficiency function to asymptotic performance predictions of given complex problems on given parallel computer in order to illustrate critical role of decomposition strategy and needed optimisation of communication complexity to developed effective PA.

LaGP: Large-Scale Spatial Modeling via Local Approximate Gaussian Processes inR

Gaussian process (GP) regression models make for powerful predictors in out of sample exercises, but cubic runtimes for dense matrix decompositions severely limit the size of data - training and testing - on which they can be deployed. That means that in computer experiment, spatial/geo-physical, and machine learning contexts, GPs no longer enjoy privileged status as data sets continue to balloon in size. We discuss an implementation of local approximate Gaussian process models, in the laGP package for R, that offers a particular sparse-matrix remedy uniquely positioned to leverage modern parallel computing architectures. The laGP approach can be seen as an update on the spatial statistical method of local kriging neighborhoods. We briefly review the method, and provide extensive illustrations of the features in the package through worked-code examples. The appendix covers custom building options for symmetric multi-processor and graphical processing units, and built-in wrapper routines that automate distribution over a simple network of workstations.

Prestack Depth Migration by a Parallel 3D PSPI

Prestack depth migration for seismic reflection data is commonly used tool for imaging complex geological structures such as salt domes, faults, thrust belts, and stratigraphic structures. Phase shift plus interpolation (PSPI) algorithm is a useful tool to directly solve a wave equation and the results have natural properties of the wave equation. Amplitude and phase characteristics, in particular, are better preserved. The PSPI algorithm is widely used in hydrocarbon exploration because of its simplicity, efficiency, and reduced efforts for computation. However, meaningful depth image of 3D subsurface requires parallel computing to handle heavy computing time and great amount of input data. We implemented a parallelized version of 3D PSPI for prestack depth migration using Open-Multi-Processing (Open MP) library. We verified its performance through applications to 3D SEG/EAGE salt model with a small scale Linux cluster. Phase-shift was performed in the vertical and horizontal directions, respectively, and then interpolated at each node. This gave a single image gather according to shot gather. After summation of each single image gather, we got a 3D stacked image in the depth domain. The numerical model example shows good agree- ment with the original geological model.

Scalability of Direct Solver for Non-stationary Cahn-Hilliard Simulations with Linearized time Integration Scheme

We study the features of a new mixed integration scheme dedicated to solving the non-stationary variational problems. The scheme is composed of the FEM approximation with respect to the space variable coupled with a 3-leveled time integration scheme with a linearized right-hand side operator. It was applied in solving the Cahn-Hilliard parabolic equation with a nonlinear, fourth-order elliptic part. The second order of the approximation along the time variable was proven. Moreover, the good scalability of the software based on this scheme was confirmed during simulations. We verify the proposed time integration scheme by monitoring the Ginzburg-Landau free energy. The numerical simulations are performed by using a parallel multi-frontal direct solver executed over STAMPEDE Linux cluster. Its scalability was compared to the results of the three direct solvers, including MUMPS, SuperLU and PaSTiX.

Event Processing Performance of KIAF Based on PROOF

Managing and processing the large scale of data is the crucial part of the experimental high energy physics which uses a gigantic instrument in order to exploit the world’s largest particle accelerator at CERN. The WLCG has enabled the distributed computing technology, which is called the grid, and has proven its excellent performance by contributing their capability of data management and processing. In addition to the grid computing, a compact but powerful computing facility introducing the parallelism is designed. In this paper, we introduce the KIAF cluster which is designed to process large scale of data produced from the ALICE experiment at CERN LHC in parallel based on PROOF. The PROOF enables parallelism on a Linux cluster by exploiting the special characteristic of the data produced in high energy physics. Event processing performance, the number of event processed per second and the size of data processed per second, of the KIAF cluster is shown with the practical use case for the high energy physics. The performance of the KIAF shows pseudo-linearity increasing as the number of workers involved in the processing while it shows as well an evidence of the upper limitation of the scalability of PROOF cluster in terms of the performance.