• Journal of Internet Computing and Services
• /
• v.2 no.3
• /
• pp.81-92
• /
• 2001

The clustering environment with heterogeneous workstations provides the cost effectiveness and usability for executing applications in parallel. The load balancing is considered as a necessary feature for the clustering of heterogeneous workstations to minimize the turnaround time. Since each workstation may have different users, groups. requests for different tasks, and different processing power, the capability of each processing unit is relative to the others' unit in the clustering environment Previous works is a static approach which assign a predetermined weight for the processing capability of each workstation or a dynamic approach which executes a benchmark program to get relative processing capability of each workstation. The execution of the benchmark program, which has nothing to do with the application being executed, consumes the computation time and the overall turnaround time is delayed. In this paper, we present an efficient task distribution method and implementation of load balancing system for the clustering environment with heterogeneous workstations. Turnaround time of the methods presented in this paper is compared with the method without load balancing as well as with the method load balancing with performance evaluation program. The experimental results show that our methods outperform all the other methods that we compared.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.05a
• /
• pp.5-7
• /
• 2023

한국과학기술정보연구원(KISTI)의 슈퍼컴퓨터 보조시스템인 Neuron은 이기종 가속기인 GPU가 탑재된 클러스터 시스템으로 작업관리 소프트웨어인 SLURM을 통해 국내 연구자들에게 서비스되고 있다. 본 논문에서는 SLURM 작업관리 소프트웨어의 작업 스케줄링 정책을 이용하여 연구자들이 제출하는 복수 개의 대기작업을 공정하게 처리하는 방안에 대해서 소개한다.

• The KIPS Transactions:PartB
• /
• v.8B no.1
• /
• pp.89-97
• /
• 2001

대용량 멀티미디어 미디어 서버를 구성함에 있어 I/O 병목현상을 극복하기 위하여 저장 서버들과 제어 서버로 구성되어진 2계층 분산 클러스터 서버구조가 많이 사용된다. 2 계층 분산 클러스터 서버는 부하 균등, 대역폭 관리 및 저장 서버의 관리 측면에서 유리한 반면, 저장 서버와 제어 서버간의 통신 오버헤드를 발생시킨다. 이러한 오버헤드를 줄이기 위해서는 저장 서버에서 읽은 미디어 데이터를 제어 서버를 거치지 않고 직접 클라이언트에 전송할 수 있어야 한다. 그리고, 저장 용량을 확장하거나 손상된 디스크를 교체하는 경우를 대비하여 분산 클러스터 서버는 다양한 성능의 이기종 디스크를 지원하여야 한다. 또한, I/O 장치와 운영체제가 빠르게 발전됨에 따라 미디어 서버는 새로운 I/O 장치 및 운영체제 등에 쉽게 이식될 수 있어야 하고, 응용 소프트웨어 개발자가 시스템의 환경에 따라 블록크기, 데이터 배치정책, 사본 정책 등을 유연하게 조절할 수 있어야 한다. 본 논문에서 위에서 언급한 멀티미디어 서버의 요구를 고려하여 Fast Ethernet 환경에서 병렬 멀티미디어 파일 시스템(PMFS : Parallel Multimedia File System)을 설계 및 구현하고 실험을 통해 PVFS(Parallel Virtual File System)와 성능을 비교 분석하였다. 이 실험의 결과에 따르면 PMFS는 멀티미디어 데이터에 대하여 PVFS보다 3%∼15%의 향상된 성능을 보였다.

• Journal of the Korea Society of Computer and Information
• /
• v.13 no.5
• /
• pp.219-227
• /
• 2008

As found in research on constraint satisfaction problems, the choice of variable ordering heuristics is crucial for effective solving of constraint optimization problems. For the special problems such as energy-efficient clustering in heterogeneous wireless sensor networks, in which cluster heads have an inclination to be near a base station, we propose a new approach based on the static preferences variable orderings and provide a pnode heuristic algorithm for a specific application. The pnode algorithm selects the next variable with the highest Preference. In our problem, the preference becomes higher when the cluster heads are closer to the optimal region, which can be obtained a Priori due to the characteristic of the problem. Since cluster heads are the most dominant sources of Power consumption in the cluster-based sensor networks, we seek to minimize energy consumption by minimizing the maximum energy dissipation at each cluster heads as well as sensor nodes. Simulation results indicate that the proposed approach is more efficient than other methods for solving constraint optimization problems with static preferences.

• The KIPS Transactions:PartA
• /
• v.14A no.2
• /
• pp.99-106
• /
• 2007

Grid computing, a mechanism which uses heterogeneous systems that are geographically distributed, draws attention as a new paradigm for the next generation operation of parallel and distributed computing. The importance of grid computing concerning communication cost is very huge because grid computing furnishes uses with integrated virtual computing service, in which a number of computer systems are connected by a high-speed network. Therefore, to reduce the execution time, the scheduling algorithm in grid environment should take communication cost into consideration as well as computing ability of resources. However, most scheduling algorithms have not only ignored the communication cost by assuming that all tasks were dealt in one cluster, but also did not consider the overhead of communication cost when the tasks were processed in a number of clusters. In this paper, the functions of original scheduling algorithms are analyzed. More importantly, the functions of algorithms are compared and analyzed with consideration of communication cost within the co allocation environment, in which a task is performed separately in many clusters.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.24 no.3
• /
• pp.63-67
• /
• 2024

Recently, artificial intelligence (AI) workloads encompassing various industries such as smart logistics, FinTech, and entertainment are being executed on the cloud. In this paper, we address the scheduling issues of various AI workloads on a multi-tenant cloud system composed of heterogeneous GPU clusters. Traditional scheduling decreases GPU utilization in such environments, degrading system performance significantly. To resolve these issues, we present a new scheduling approach utilizing genetic algorithm-based optimization techniques, implemented within a process-based event simulation framework. Trace driven simulations with diverse AI workload traces collected from Alibaba's MLaaS cluster demonstrate that the proposed scheduling improves GPU utilization compared to conventional scheduling significantly.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2016.05a
• /
• pp.376-377
• /
• 2016

Performance of Hadoop, which is a distributed data processing framework for big data analysis, is affected by several characteristics of each node in distributed cluster such as processing power and network bandwidth. This paper analyzes previous approaches for heterogeneous hadoop clusters, and presents several requirements for distributed node clustering in small-and-medium sized hospitals by considering computing environments of the hospitals.

• KIISE Transactions on Computing Practices
• /
• v.22 no.12
• /
• pp.663-674
• /
• 2016

Traditional supercomputing services suffer from limited accessibility and low utilization in that users(researchers) may perform computational executions only using terminal-based command line interfaces. To address this problem, in this paper, we provide the design and implementation details of National supercomputing service framework. The proposed framework supports all the fundamental primitive functions such as user management/authentication, heterogeneous computing resource management, HPC (High Performance Computing) job management, etc. so that it enables various 3rd-party applications to be newly built on top of the proposed framework. Our framework also provides Web-based RESTful OpenAPIs and the abstraction interfaces of job schedulers (as well as bundle scheduler plug-ins, for example, LoadLeveler, Open Grid Scheduler, TORQUE) in order to easily integrate the broad spectrum of heterogeneous computing clusters. To show and validate the effectiveness of the proposed framework, we describe the best practice scenario of high energy physics Lattice-QCD as an example application.

• Journal of Digital Convergence
• /
• v.19 no.3
• /
• pp.211-216
• /
• 2021

Information from various heterogeneous devices is steadily increasing in distributed cloud environments. This is because high-speed network speeds and high-capacity multimedia data are being used. However, research is still underway on how to minimize information errors in big data sent and received by heterogeneous devices. In this paper, we propose a deep learning-based asymmetric storage management technique for minimizing bandwidth and data errors in networks generated by information sent and received in cloud environments. The proposed technique applies deep learning techniques to optimize the load balance after asymmetric hash of the big data information generated by each device. The proposed technique is characterized by allowing errors in big data collected from each device, while also ensuring the connectivity of big data by grouping big data into groups of clusters of dogs. In particular, the proposed technique minimizes information errors when storing and managing big data asymmetrically because it used a loss function that extracted similar values between big data as seeds.

• Journal of KIISE
• /
• v.42 no.5
• /
• pp.573-579
• /
• 2015

Nationwide supercomputing infrastructures in Korea consist of geographically distributed supercomputing clusters. We developed High-Throughput Computing as a Service(HTCaaS) based on these distributed national supecomputing clusters to facilitate the ease at which scientists can explore large-scale and complex scientific problems. In this paper, we present our mechanism for dynamically managing computing resources and show its effectiveness through a case study of a real scientific application called drug repositioning. Specifically, we show that the resource utilization, accuracy, reliability, and usability can be improved by applying our resource management mechanism. The mechanism is based on the concepts of waiting time and success rate in order to identify valid computing resources. The results show a reduction in the total job completion time and improvement of the overall system throughput.