• Journal of KIISE:Computing Practices and Letters
• /
• v.5 no.6
• /
• pp.700-707
• /
• 1999

본 논문에서는 GIS 관련 연산을 실시간에 효율적으로 처리하기 위한 분산공유메모리 기반 병렬처리 시스템을 제안한다. 본 논문의 분산공유메모리 시스템은 메시지전달 방식의 분산메모리 MIMD 컴퓨터 상에 소프트웨어 기반 분산공유메모리 모듈을 탑재함으로써 구현되었다. 또한 GIS 연산의 기본이 되는 공간 객체를 공유의 기본 단위로 설정하고, GIS 데이타의 특성을 반영하여 읽기전용 공유데이타 타입을 추가하였으며, 네트워크 오버헤드를 줄이기 위하여 복수의 객체를 한번에 읽어오는 bulk access가 가능하도록 하였다. 본 시스템에서는 GIS 데이타의 효율적인 분배를 위하여 부하균등화 기법으로 guided self scheduling을 사용하였다. 실험결과 본 시스템은 네트워크 캐쉬의 효율적인 활용을 통하여 소프트웨어 기반 분산메모리 시스템의 오버헤드에도 불구하고 MPI 기반 메시지전달 방식에 비하여 향상된 성능을 얻을 수 있었다.Abstract In this paper, we propose a distributed shared memory(DSM) based parallel processing system to process GIS related computations efficiently in real time. The system is based on a software DSM module implemented on top of a distributed MIMD computer. In the DSM system, spatial object, which is a fundamental structure to represent GIS data, is used as a basic unit for sharing, and a read-only shared data type is added to reflect the characteristics of GIS data. In addition, a bulk access to multiple shared data is made possible to reduce the network overhead. A guided self scheduling method is devised for efficient load balancing in distributing GIS data to parallel processors. The experimental results show that the DSM system performs better than an MPI based message-passing system through the efficient utilization of network cache in spite of the system's software overhead.

• Journal of the Korean Data and Information Science Society
• /
• v.26 no.5
• /
• pp.1035-1045
• /
• 2015

Recent studies in Big Data Analysis are showing promising results, utilizing the main memory for rapid data processing. In-memory computing technology can be highly advantageous when used with high-performing servers having tens of gigabytes of RAM with multi-core processors. The constraint in network in these infrastructure can be lessen by combining in-memory technology with distributed parallel processing. This paper discusses the research in the aforementioned concept applying to a test taxi hailing application without disregard to its underlying RDBMS structure. The application of IMDG technology in the application's backend API without restructuring the database schema yields 6 to 9 times increase in performance in data processing and throughput. Specifically, the change in throughput is very small even with increase in data load processing.

• Journal of Electrical Engineering and Technology
• /
• v.6 no.4
• /
• pp.573-579
• /
• 2011

To meet the scalability and performance requirements of data analyses, which often involve voluminous data, efficient parallel or concurrent algorithms and frameworks are essential. We present a high-performance Korean morphological analyzer which employs the MapReduce framework on the graphics processing unit (GPU). MapReduce is a programming framework introduced by Google to aid the development of web search applications on a large number of central processing units (CPUs). GPUs are designed as a special-purpose co-processor. Their programming interfaces are typically formulated for graphics applications. Compared to CPUs, GPUs have greater computation power and memory bandwidth; however, GPUs are more difficult to program because of the design of their architectures. The performance of the Korean morphological analyzer using the MapReduce framework on the GPU is evaluated in comparison with the CPU-based model. The proposed Korean Morphological analyzer shows promising scalable performance on distributed computing with the GPU.

• Journal of KIISE
• /
• v.43 no.3
• /
• pp.304-313
• /
• 2016

We present a novel memory-efficient parallel ray casting algorithm for unstructured grid volume rendering on multi-core CPUs. Our method is based on the Bunyk ray casting algorithm. To solve the high memory overhead problem of the Bunyk algorithm, we allocate a fixed size local buffer for each thread and the local buffers contain information of recently visited faces. The stored information is used by other rays or replaced by other face's information. To improve the utilization of local buffers, we propose an image-plane based ray grouping algorithm that makes ray groups have high coherency. The ray groups are then distributed to computing threads and each thread processes the given groups independently. We also propose a novel hash function that uses the index of faces as keys for calculating the buffer index each face will use to store the information. To see the benefits of our method, we applied it to three unstructured grid datasets with different sizes and measured the performance. We found that our method requires just 6% of the memory space compared with the Bunyk algorithm for storing face information. Also it shows compatible performance with the Bunyk algorithm even though it uses less memory. In addition, our method achieves up to 22% higher performance for a large-scale unstructured grid dataset with less memory than Bunyk algorithm. These results show the robustness and efficiency of our method and it demonstrates that our method is suitable to volume rendering for a large-scale unstructured grid dataset.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.25 no.5B
• /
• pp.899-912
• /
• 2000

PVM provides users with a single image of high performance parallel computing machine by collecting machines distributed over a network. Low communication overhead is essential to effectively run applications on PVM based platforms. In the original PVM, three times of memory copies are required for a PVM task to send a message to a remote task, which results in performance degradation. We propose a zero-copy model using global shared memory that can be accessed by PVM tasks, PVM daemon, and network interface card(NIC). In the scheme, a task packs data into global shared memory, and notify daemon that the data is ready to be sent, then daemon routes the data to a remote task to which it is sent with no virtual data copy overhead. Experimental result reveals that the message round trip time between two machines is reduced significantly in the proposed zero-copy scheme.

• ETRI Journal
• /
• v.42 no.5
• /
• pp.748-760
• /
• 2020

MapReduce is a framework that can process huge datasets in parallel and distributed computing environments. However, a single machine failure during the runtime of MapReduce tasks can increase completion time by 50%. MapReduce handles task failures by restarting the failed task and re-computing all input data from scratch, regardless of how much data had already been processed. To solve this issue, we need the computed key-value pairs to persist in a storage system to avoid re-computing them during the restarting process. In this paper, the task failure resilience (TFR) technique is proposed, which allows the execution of a failed task to continue from the point it was interrupted without having to redo all the work. Amazon ElastiCache for Redis is used as a non-volatile cache for the key-value pairs. We measured the performance of TFR by running different Hadoop benchmarking suites. TFR was implemented using the Hadoop software framework, and the experimental results showed significant performance improvements when compared with the performance of the default Hadoop implementation.

• KIISE Transactions on Computing Practices
• /
• v.24 no.3
• /
• pp.105-112
• /
• 2018

As data size increase, a single database is not enough to serve current volume of tasks. Since data is partitioned and stored into multiple databases, analysis should also support parallelism in order to increase efficiency. However, traditional analysis requires data to be transferred out of database into nodes where analytic service is performed and user is required to know both database and analytic framework. In this paper, we propose an efficient way to perform K-means clustering algorithm inside the distributed column-based database and relational database. We also suggest an efficient way to optimize K-means algorithm within relational database.

• The Transactions of the Korea Information Processing Society
• /
• v.6 no.9
• /
• pp.2311-2319
• /
• 1999

We present two heuristic algorithms for the task allocation problem (NP-complete problem) in parallel computing. The problem is to find an optimal mapping of multiple communicating tasks of a parallel program onto the multiple processing nodes of a distributed-memory multicomputer. The purpose of mapping these tasks into the nodes of the target architecture is the minimization of parallel execution time without sacrificing solution quality. Many heuristic approaches have been employed to obtain satisfactory mapping. Our heuristics are based on genetic algorithms and simulated annealing. We formulate an objective function as a total computational cost for a mapping configuration, and evaluate the performance of our heuristic algorithms. We compare the quality of solutions and times derived by the random, greedy, genetic, and annealing algorithms. Our experimental findings from a simulation study of the allocation algorithms are presented.

• Journal of KIISE:Computing Practices and Letters
• /
• v.6 no.5
• /
• pp.457-467
• /
• 2000

본 논문에서는 분산 메모리 구조를 갖는 병렬 컴퓨터 상에서 방대한 크기를 갖는 볼륨 데이터의 효과적인 가시화를 위한 병렬 광선 투사법을 제안한다. 데이터의 압축을 기반으로 하는 본 기법은 다른 프로세서의 메모리로부터 데이터를 읽기보다는 자신의 지역 메모리에 존재하는 압축된 데이터를 빠르게 복원함으로써 병렬 렌더링 성능을 향상시키는 것을 목표로 한다. 본 기법은 객체-순서와 영상-순서 탐색 알고리즘 모두의 정점을 이용하여 성능을 향상시켰다. 즉, 블록 단위의 최대-최소 팔진트리의 탐색과 각 픽셀의 불투명도 값을 동적으로 유지하는 실시간 사진트리를 응용함으로써 객체-공간과 영상-공간 각각의 응집성을 이용하였다. 본 논문에서 제안하는 압축 기반 병렬 볼륨 렌더링 방법은 렌더링 수행 중 발생하는 프로세서간의 통신을 최소화하도록 구현되었는데, 이러한 특징은 프로세서 사이의 상당히 높은 데이터 통신 비용을 감수하여야 하는 PC 및 워크스테이션의 클러스터와 같은 더욱 실용적인 분산 환경에서 매우 유용하다. 본 논문에서는 Cray T3E 병렬 컴퓨터 상에서 Visible Man 데이터를 이용하여 실험을 수행하였다.

• KIPS Transactions on Computer and Communication Systems
• /
• v.7 no.4
• /
• pp.99-102
• /
• 2018

Caffe is a neural net learning software which is widely used in academic researches. The GPU memory capacity is one of the most important aspects of designing neural net architectures. For example, many object detection systems require to use less than 12GB to fit a single GPU. In this paper, we extended Caffe to allow to use more than 12GB GPU memory. To verify the effectiveness of the extended software, we executed some training experiments to determine the learning efficiency of the object detection neural net software using a PC with three GPUs.