• Journal of the Korean Operations Research and Management Science Society
• /
• v.39 no.4
• /
• pp.19-31
• /
• 2014

We study parallel processing techniques for the R programming language of high performance computing technology. In this study, we used massively parallel computing system which has 25,408 cpu cores. We conducted a performance evaluation of a distributed memory system using MPI and of a the shared memory system using OpenMP. Our findings are summarized as follows. First, For some particular algorithms, parallel processing is about 150 times faster than serial processing in R. Second, the distributed memory system gets faster as the number of nodes increases while shared memory system is limited in the improvement of performance, due to the limit of the number of cpus in a single system.

• ETRI Journal
• /
• v.43 no.5
• /
• pp.825-834
• /
• 2021

Based on its simplicity and user-friendly characteristics, OpenMP has become the standard model for programming on shared-memory architectures. Checkpointing-aided parallel execution (CAPE) is an approach that utilizes the discontinuous incremental checkpointing technique (DICKPT) to translate and execute OpenMP programs on distributed-memory architectures automatically. Currently, CAPE implements the OpenMP execution model by utilizing the DICKPT to distribute parallel jobs and their data to slave machines, and then collects the results after executing these distributed jobs. Although this model has been proven to be effective in terms of performance and compatibility with OpenMP on distributed-memory systems, it cannot fully exploit the capabilities of multicore processors. This paper presents a novel execution model for CAPE that utilizes two levels of parallelism. In the proposed model, we add another level of parallelism in the form of multithreaded processes on slave machines with the goal of better exploiting their multicore CPUs. Initial experimental results presented near the end of this paper demonstrate that this model provides significantly enhanced CAPE performance.

• Journal of Communications and Networks
• /
• v.16 no.4
• /
• pp.436-446
• /
• 2014

The high intensity of research and modeling in fields of mathematics, physics, biology and chemistry requires new computing resources. For the big computational complexity of such tasks computing time is large and costly. The most efficient way to increase efficiency is to adopt parallel principles. Purpose of this paper is to present the issue of parallel computing with emphasis on the analysis of parallel systems, the impact of communication delays on their efficiency and on overall execution time. Paper focuses is on finite algorithms for solving systems of linear equations, namely the matrix manipulation (Gauss elimination method, GEM). Algorithms are designed for architectures with shared memory (open multiprocessing, openMP), distributed-memory (message passing interface, MPI) and for their combination (MPI + openMP). The properties of the algorithms were analytically determined and they were experimentally verified. The conclusions are drawn for theory and practice.

• Korean Journal of Computational Design and Engineering
• /
• v.16 no.3
• /
• pp.236-245
• /
• 2011

Many researchers have recently studied multi-level formulation strategies to solve the MDO problems and they basically distributed the coupling compatibilities across all disciplines, while single-level formulations concentrate all the controls at the system-level. In addition, approximation techniques became remedies for computationally expensive analyses and simulations. This paper studies comparisons of the MDO methods with respect to computing performance considering both conventional sequential and modem distributed/parallel processing environments. The comparisons show Individual Disciplinary Feasible (IDF) formulation is the most efficient for sequential processing and IDF with approximation (IDFa) is the most efficient for parallel processing. Results incorporating to popular design examples show this finding. The author suggests design engineers should firstly choose IDF formulation to solve MDO problems because of its simplicity of implementation and not-bad performance. A single drawback of IDF is requiring more memory for local design variables and coupling variables. Adding cheap memories can save engineers valuable time and effort for complicated multi-level formulations and let them free out of no solution headache of Multi-Disciplinary Analysis (MDA) of the Multi-Disciplinary Feasible (MDF) formulation.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.41 no.11
• /
• pp.1490-1501
• /
• 2016

In this paper, we propose an image-based indoor localization system using parallel distributed computing. In order to reduce computation time for indoor localization, an scale invariant feature transform (SIFT) algorithm is performed in parallel by using Apache Spark. Toward this goal, we propose a novel image processing interface of Apache Spark. The experimental results show that the speed of the proposed system is about 3.6 times better than that of the conventional system.

• Journal of Internet Computing and Services
• /
• v.18 no.2
• /
• pp.53-60
• /
• 2017

Since processors are handling inputs and outputs independently on distributed memory computers, different file input/output methods are used. In this paper, we implemented and compared various file I/O methods to show their efficiency on distributed memory parallel computers. The implemented I/O systems are as following: (i) parallel I/O using NFS, (ii) sequential I/O on the host processor and domain decomposition, (iii) MPI-IO. For performance analysis, we used a separated file server and multiple processors on one or two computational servers. The results show the file I/O with NFS for inputs and sequential output with domain composition for outputs are best efficient respectively. The MPI-IO result shows unexpectedly the lowest performance.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.15 no.5
• /
• pp.1649-1665
• /
• 2021

In view of the low accuracy of the traditional FunkSVD algorithm, and in order to improve the computational efficiency of the algorithm, this paper proposes a parallel algorithm of improved FunkSVD based on Spark (SP-FD). Using RMSProp algorithm to improve the traditional FunkSVD algorithm. The improved FunkSVD algorithm can not only solve the problem of decreased accuracy caused by iterative oscillations but also alleviate the impact of data sparseness on the accuracy of the algorithm, thereby achieving the effect of improving the accuracy of the algorithm. And using the Spark big data computing framework to realize the parallelization of the improved algorithm, to use RDD for iterative calculation, and to store calculation data in the iterative process in distributed memory to speed up the iteration. The Cartesian product operation in the improved FunkSVD algorithm is divided into blocks to realize parallel calculation, thereby improving the calculation speed of the algorithm. Experiments on three standard data sets in terms of accuracy, execution time, and speedup show that the SP-FD algorithm not only improves the recommendation accuracy, shortens the calculation interval compared to the traditional FunkSVD and several other algorithms but also shows good parallel performance in a cluster environment with multiple nodes. The analysis of experimental results shows that the SP-FD algorithm improves the accuracy and parallel computing capability of the algorithm, which is better than the traditional FunkSVD algorithm.

• Journal of KIISE:Computing Practices and Letters
• /
• v.10 no.1
• /
• pp.49-56
• /
• 2004

Load imbalance is a serious impediment to achieving good performance in parallel processing. Global load balancing schemes cannot adequately manage to balance parallel tasks generated from a single application. Dynamic loop scheduling methods are known to be useful in balancing parallel loops on shared-memory multiprocessor machines. However, their centralized nature causes a bottleneck for the relatively small number of processors in a network of workstations because of order-of-magniture differences in communication overheads. Moreover, improvements of basis loops scheduling methods have not effectively dealt with irregularly distributed workloads in parallel loops, which commonly occur in applications for a network of workstation. In this paper, we present a new reconfigurable and decentralized balancing method for parallel loops on a network of workstations. Since our method supplements performance balancing with those tranditional load balancing methods, it minimizes the overall execution time.

• Journal of KIISE
• /
• v.42 no.8
• /
• pp.998-1009
• /
• 2015

In recent years, there has been a growing interest in RDFS Inference to build a rich knowledge base. However, it is difficult to improve the inference performance with large data by using a single machine. Therefore, researchers are investigating the development of a RDFS inference engine for a distributed computing environment. However, the existing inference engines cannot process data in real-time, are difficult to implement, and are vulnerable to repetitive tasks. In order to overcome these problems, we propose a method to construct an in-memory distributed inference engine that uses a parallel graph structure. In general, the ontology based on a triple structure possesses a graph structure. Thus, it is intuitive to design a graph structure-based inference engine. Moreover, the RDFS inference rule can be implemented by utilizing the operator of the graph structure, and we can thus design the inference engine according to the graph structure, and not the structure of the data table. In this study, we evaluate the proposed inference engine by using the LUBM1000 and LUBM3000 data to test the speed of the inference. The results of our experiment indicate that the proposed in-memory distributed inference engine achieved a performance of about 10 times faster than an in-storage inference engine.

• Nuclear Medicine and Molecular Imaging
• /
• v.43 no.5
• /
• pp.443-450
• /
• 2009

Purpose: Conventional image reconstruction uses simplified physical models of projection. However, real physics, for example 3D reconstruction, takes too long time to process all the data in clinic and is unable in a common reconstruction machine because of the large memory for complex physical models. We suggest the realistic distributed memory model of fast-reconstruction using parallel processing on personal computers to enable large-scale technologies. Materials and Methods: The preliminary tests for the possibility on virtual manchines and various performance test on commercial super computer, Tachyon were performed. Expectation maximization algorithm with common 2D projection and realistic 3D line of response were tested. Since the process time was getting slower (max 6 times) after a certain iteration, optimization for compiler was performed to maximize the efficiency of parallelization. Results: Parallel processing of a program on multiple computers was available on Linux with MPICH and NFS. We verified that differences between parallel processed image and single processed image at the same iterations were under the significant digits of floating point number, about 6 bit. Double processors showed good efficiency (1.96 times) of parallel computing. Delay phenomenon was solved by vectorization method using SSE. Conclusion: Through the study, realistic parallel computing system in clinic was established to be able to reconstruct by plenty of memory using the realistic physical models which was impossible to simplify.