Earticle

최근 유전체 분석 기법과 시퀀싱 기기의 발전에 따라 개인별 맞춤 의학의 실현에 관한 기대가 높아지고 있다. 개인 맞춤 의학은 개인의 유전 정보를 분석하고 이 들의 특이성을 파악하여, 개인의 질병 예방 및 치료에 적용하고자 하는 전략이다. 그러나 한 개인의 시퀀스 분석을 위한 적정 커버리지(약 30 커버리지로 예측됨)의 시퀀싱 데이터는 100 GB를 넘는 방대한 크기를 가지므로 다수 개인의 유전 정보의 분석/비교를 위하여 대규모 데이터 처리/분석 환경 및 고성능 컴퓨팅 방식의 도입이 필수적이다. 본 연구에서는 클라우드 컴퓨팅 기술을 기반으로 하는 새로운 병렬 유전체 단위반복변위 (Copy Number Variation, CNV) 영역 추출 알고리즘을 제안한다. 제안된 방식에서는 다수의 컴퓨팅 노드를 활용하는 병렬 처리를 위하여 Hadoop의 MapReduce 오픈 소스를 사용하며, 대규모 유전자 시퀀스를 병렬로 처리하여 (개인_ID, 염색체_ID, 컨티그_ID) 단위로 CNV 영역을 추출, 보고한다.

Recent advances in genome analysis method and sequencing technology may allow for a greater degree of personalized medicine than is currently available. Personalized medicine is a future medical model emphasizing the systematic use of genetic information about an individual patient to apply and optimize patient's preventative and therapeutic care. However, as an enormous amount of data (around 100 GB or up to 30x read coverage data for an individual) is needed to analyze and compare genetic information among many individuals, it is essential to employ the environment of high-performance computing system. In this study, we propose a novel algorithm to detect CNV (Copy Number Variation) regions by carrying out parallel and transparency processing based on cloud computing in a part of new frontier technologies. This method is able to perform simultaneous tasking with large numbers of computing nodes using open source of 'MapReduce' in 'Hadoop' project. The CNV regions detected by the proposed method are reported as a unit of map consists of <individual_id>, <chromosome_id>, and <contig_id> by carrying out parallel processing for tremendous size of short read data from next generation sequencing instrument.

최근 클라우드 컴퓨팅이 발전함에 따라 데이터베이스 아웃소싱에 대한 관심이 증가하고 있다. 그러나 데이터베이스 를 아웃소싱하는 경우 데이터 소유자의 민감한 데이터가 노출될 수 있다는 문제점이 존재한다. 따라서 본 논문에서 는 프라이버시 보호를 지원하는 병렬 kNN 분류 알고리즘을 제안한다. 제안하는 알고리즘은 기존 기법과 동일한 수 준의 정보 보호 수준을 제공하면서 다중 CPU 코어를 활용한 병렬 처리를 지원한다. 이를 위해 제안하는 알고리즘 은 단일 CPU 코어 기반 데이터 처리 프로토콜을 다중 CPU 코어 기반 데이터 처리 프로토콜로 변환함으로써 효율 적인 질의처리를 수행한다. 또한 노이즈(Noise) 데이터를 전처리함으로써 정보 보호를 지원하는 동시에 효율적인 질의 처리가 가능하다. 마지막으로 성능 평가를 통해 제안하는 알고리즘이 기존 기법보다 질의처리 시간 측면에서 8~30배 성능이 우수함을 보인다.

With the recent development of cloud computing, interest in database outsourcing is increasing. However, when outsourcing a database, there is a problem that sensitive data of the data owner may be exposed. Therefore, in this paper, we propose a parallel kNN classification algorithm that supports information protection. The proposed algorithm supports parallel processing using multiple CPU cores while providing the same level of information protection as the existing techniques. For this, the proposed algorithm performs efficient query processing by converting a data processing protocol based on a single CPU core into a data processing protocol based on multiple CPU cores. By preprocessing noise data, it also supports information protection and enables efficient query processing. Finally, it is shown from performance evaluation that the proposed algorithm is 8 to 30 times better than the existing techniques, in terms of query processing time.

다양한 병렬 컴퓨팅 시스템을 지원하기 위해서는 LLVM IR을 벡터/행렬을 보다 효과적으로 지원할 수 있도록 확장하는 것과 LLVM IR을 machine code로 바꾸어 주는 부분을 새로운 알고리즘으로 설계하여 구현하면 된다. IR 예제에서 보았듯이 기본적으로 RISC 명령어로 구성되어 있기 때문에 RISC 명령어 생성은 자연스럽게 생성되며, 벡 터 또한 현재 지원가능한데 행렬 명령어는 지원되지 못하고 있다. 벡터/행렬을 보다 강력하게 지원하기 위한 새로운 IR 구조, 명령어 생성 알고리즘 및 관련 부분의 확장이 필요하다. 이를 위해 LLVM IR의 각 명령어를 (벡터/행렬을 위한) target architecture의 적당한 명령어로 mapping을 해주는 부분 (instruction selection 알고리즘)이 중요하다. LLVM IR 명령어의 의미를 파악하고, target architecture의 각 명령어 의미와 syntax를 비교하여, 패턴이 일치하는 명령어를 선택하여 mapping을 효율적으로 해줘야 한다.

In order to support various parallel computing systems, it is necessary to extend LLVM IR to more efficiently support vector / matrix and to design LLVM IR to machine code as a new algorithm. As shown in the IR example, RISC instruction generation is naturally generated because the RISC instruction is basically composed of the RISC instruction, and the vector instruction is also not supported. There is a need for new IR structures, command generation algorithms and related extensions to support vector / matrix more robustly. To do this, it is important to map each instruction in the LLVM IR to the appropriate instruction in the target architecture (vector / matrix) (instruction selection algorithm). It is necessary to understand the meaning of LLVM IR command, to compare the meaning of each instruction of the target architecture with syntax, and to select the instruction that matches the pattern to make mapping efficient.

As the performance of the system increases, more parallelized data is being processed than single processing of data. Today's cpu structure has been developed to leverage multicore, and hence data processing methods are being developed to enable parallel processing. In recent years desktop cpu has increased multicore, data is growing exponentially, and there is also a growing need for data processing as artificial intelligence develops. This neural network of artificial intelligence consists of a matrix, making it advantageous for parallel processing. This paper aims to speed up the processing of the system by using raspberrypi to implement the cluster building and parallel processing system against the backdrop of the foregoing discussion. Raspberrypi is a credit card-sized single computer made by the raspberrypi Foundation in England, developed for education in schools and developing countries. It is cheap and easy to get the information you need because many people use it. Distributed processing systems should be supported by programs that connected multiple computers in parallel and operate on a built-in system. RaspberryPi is connected to switchhub, each connected raspberrypi communicates using the internal network, and internally implements parallel processing using the Message Passing Interface (MPI). Parallel processing programs can be programmed in python and can also use C or Fortran. The system was tested for parallel processing as a result of multiplying the two-dimensional arrangement of 10000 size by 0.1. Tests have shown a reduction in computational time and that parallelism can be reduced to the maximum number of cores in the system. The systems in this paper are manufactured on a Linux-based single computer and are thought to require testing on systems in different environments.

With the rapid development of embedded system and Internet of Things technology, embedded system and smart device based on embedded system is collecting huge amounts of data, and corresponding data processing and application method have been greatly changed, different from the traditional big data and cloud computing focus , local processing and application become important trend. This paper tries to take Cortex-A7 as the main system with the stronger ability , Hadoop distributed computing system is deployed in embedded system and could meet the demand of the future data process with directly managing the resource-constrained sensors, embedded systems and smart devices . Successfully deploying over 20GB data in the test, the system is verified that it can complete most of the functions of data processing cluster , and can also manage the collected sensors and embedded system terminal ,with better research and market promotional value.

Mining association rules is an important research direction in the field of data mining. Related studies have proposed many used to efficiently find large-scale database association rules algorithm, but the research on maintenance problem of association rules is less. Especially many transaction database is always in constant updates. Increase or decrease occurs when the database or dataset minimum support after the change, how to maintain the association rules have been, it got the attention of many researchers. Based on IFP-Growth increment of association rules mining model and to modify the FP-tree, put forward the suitable for transaction data and support the tree model of change, at the same time under different conditions is given incremental association rules mining algorithm, and reduce the frequency of the original dataset range query and query, and in a case of massive dataset multi-level tree structure decomposition, dynamic allocation rule tree branches, ensure load balancing, improve operation efficiency.

Multicore has long been considered an attractive platform for string matching. However, some existing traditional algorithms of string matching do not adapt to multicore platform, which pose new challenges to parallelism designs. In this paper, we introduce a multicore architecture with message passing interface to address these challenges. We exploit the popular Aho-Corasick algorithm for the string matching engine. Data parallelism is utilized to design optimization technique of string matching. The experiments show that an implementation of the 8-core system achieves up to 10.5 Gbps throughput on the average.

With the increasing number of cores in multicore processors, it is challenging task how to take advantage of powerful parallel computing for the deep packet inspection. This paper introduces the deep packet inspection with a parallel method which exploits the message-passing interface (MPI). The parallel procedure includes the master thread and the slave thread. The master assigns the data packet to the slave. The slave executes the string matching with rules for inspecting. Both the master and the slave communicate by using MPI functions. The experimental results show that the parallel method is suitable for the trend of the increasing number of cores in multicore processors. Moreover, when the number of threads is equal to the number of cores in multicore processors, the performance arrives at the maximum throughput.

Based on the alternating group strategies, a new high accurate Alternating Group 8-point scheme (AG-8p) for the diffusion problem is presented. The scheme is proved to be unconditionally stable. In numerical experiments, we compare the AG-8p scheme with the AGE and the ASC-N schemes. Finally, numerical results show that the AG-8p scheme has high accuracy.

The ultrasound image sequences are not only majorly contaminated by multiplicative noises but they are also usually contaminated with additive noises. As in the past few decades, there were some works, which had focused on removing the noises from ultrasound images, such as in the JY model [1] and in the variational model, which were able to remove both types of noises. However, denoising these noises from the ultrasound image sequence is a time-consuming process that occurred from using fixed-point iterative method. From our investigation, the most time-consuming process part of the denoising process is the Gauss-Seidel. By parallelizing these processes with modern multi-core and many-core processor, the denoising ultrasound image in real-time is possible. In this study, we propose the modified strategy from [2] for managing threads and propose the modified state-based Gauss-Seidel method from [16] for GPUs. Our proposed model can retain the frame order, and get the satisfactory frame rate (about 23.33 fps). The proposed strategy boosts the speedup of the frame denoising to 13.80 times compare to the sequential computation.

The optimization of scheduling problems is based on different criteria to optimize. One of the most important criteria is the minimization of completion time of the last task on the last machine called makespan. In this paper, we present a parallel algorithm for solving the permutation flow shop problem. Our algorithm is a basic parallel distributed algorithm deployed in a grid of computer (Grid’5000). The objective of this work is minimizing the total makespan of the tasks. Our algorithm uses the exact Branch and Bound method to find optimal solutions of the problem through the distribution of the tasks among the available processors. Computational results of our parallel algorithm using well known Taillard’s benchmarks, showed encouraging results. In particular, we succeeded to solve two new instances to optimality which had never been resolved before neither in sequential nor in parallel [29].

In the big data era, with the parallel evolution of computer architecture, computing changes and modifications of industrial application mode resource expansion capability, we need to explore a new parallel computing model, to reflect the properties and large data applications form the current parallel machines, and a variety of mainstream big data processing system for unified theoretical analysis to guide large data applications tuning. Currently, despite the large data programming model study made many achievements, and is widely used in the TB level or even PB-class data processing and analysis, but the corresponding computational model study has just begun. From traditional parallel computing model, research big data programming model and large data computation model, summed up the three basic problems of large data model, in theory, need to be addressed: the three elements of the problem model, scalability and fault tolerance issues and performance optimization. Around these three questions, on the one hand and performance optimization model to calculate the theoretical study of data from a large, on the other hand these performance optimization methods in case of an actual big data.

The demands of data analysis and processing make the parallel computing platforms are continuously developed. But the existing parallel computing models which are the core of platforms present the characteristics of diversification, high pertinence and short cycle. So a synthetic model of supporting flexible platforms urgently needs to be researched. This work researches a synthetic model to shield the heterogeneity of parallel computing models under the theories of λ-calculus, functional language and graph reduction. Based on the MapReduce and BSP models, first of all, the performing principles of models are analyzed. And then the unified modalities of models are found. Finally, the synthetic model having high performance is developed with graph reduction rules, and the experiment results are also shown.

This research contributes to the on-going research effort by exploring alternate methods for soft computing optimization the highly nonlinear and uncertain systems. This research addresses two basic issues related to the control of an uncertain system; (1) design of a robust feedback controller, and (2) the design of a parallel artificial intelligence based optimization to increase the result qualification. The robust backstepping controller proposed in this research is used to further demonstrate the appealing features exhibited by the continuum robot. Robust feedback controller is used to position control of continuum robot in presence of uncertainties. Using Lyapunov type stability arguments, a robust backstepping controller is designed to achieve this objective. The controller developed in this research is designed in two steps. Firstly, a robust stabilizing torque is designed for the nominal continuum robot dynamics derived using the constrained Lagrangian formulation. Next, the fuzzy logic methodology applied to it to solution uncertainty problem by parallel optimization. The fuzzy model free optimization is formulated to minimize the problem of nonlinear formulation of uncertain systems.

With the rapid development of the 3G network, traditional calculation methods are unable to adapt to the data scene that telecom users' Network access behavior's data scale increase rapidly dozens of TB. The cloud techniques such as Hadoop platform are introduced to solve the data storage problem. The appropriate data mining algorithms are designed from the perspective of practical application. This paper improves the traditional decision tree SPRINT algorithms, proposes a parallel computing program and successfully applies to the Hadoop platform.

In social media services and social network services, it is necessary to collect, analyze and process their big data with low maintenance cost. Therefore, distributed-parallel data processing on cloud platform is getting spotlight as useful solution for them. In this paper, we present a new architecture of DIVE-C: DIstributed-parallel Virtual Environment on Cloud computing platform for distributed parallel data processing applications which offers a transparent virtual computing environment in order to provide a way easy to launch user’s distributed parallel applications. It hides the complexity of the cloud, and helps users to focus on their new applications and core services. DIVE-C uses agent-based resource management scheme to configure VM resources and application deployment for offering various distributed-parallel application models. VM resources are automatically provided by unified cloud management layer. Furthermore, an easy-to-use web interface of DIVE-C offers convenience to users. We implemented a prototype of DIVE-C, and its experiment results show the competitive performance of DIVE-C for dynamic resource and virtual computing environment provisioning for various data processing models.

E-commerce is a cyber, electronic and informational business activities. Electronic commerce has been developing rapidly in recent years in China. Cloud computing provides a new idea for the development of electronic commerce. In practice, cloud computing also has a very good combination of e-commerce. To predict the profit of electronic commerce can find the existing problems, grasp the development trend, and make better management of electronic commerce. In this paper, we study the profit prediction of electronic commerce. Then, we propose an improved parallel PSO-LSSVM algorithm, and use this algorithm to predict the benefits of electronic commerce. Experimental results show that the proposed algorithm is effective and reliable.