• The KIPS Transactions:PartA
• /
• v.10A no.4
• /
• pp.285-294
• /
• 2003

The request scheduling algorithms being used for the cluster web servers are mostly in two categories : load-balancing and contents-based cache affinity The goal of the load-balancing algorithms is to balance the loads between real servers. On the other hand, contents-based scheduling algorithm exploits the cache affinity in a way that the same type of requests are to be directed to a dedicated real server allowing load imbalance. So the performance comparison of the two algorithms is necessary, nevertheless the related experiment results are not much suggested. In this paper, performance evaluations have been done to compare the performance of the two scheduling algorithms. To accomplish this, we first implement a linux cluster web server, and then present the performance measurement results. The main contribution of this paper is to help the cluster web server administrators to select an algorithm fitting in with their circumstances from the two algorithms.

• Proceedings of the Korean Information Science Society Conference
• /
• 2002.04a
• /
• pp.85-87
• /
• 2002

본 논문에서는 리눅스 클러스터에서 커널 수준의 응용 프로그램 개발을 위한 통신 모듈인 KCCM(Kernel level Cluster Communication Module)에 대한 구조를 기술한다. KCCM은 비동기 통신을 지원하기 위한 응용 프로그램 인터페이스와 송수신(Send/ Receive)형태의 동기 통신을 지원하기 위한 응용 프로그램 인터페이스를 함께 제공하며, 다른 플랫폼으로의 포팅을 고려하여 소켓 인터페이스를 이용해 구현되었다. 또한 장애 상황에서도 서비스를 계속하고 장애를 복구할 수 있도록 설계되어 커널 수준의 고 가용성 클러스터 응용 프로그램을 개발하는데 유용하다.

• Proceedings of the Korean Information Science Society Conference
• /
• 2006.10a
• /
• pp.490-493
• /
• 2006

본 논문은 Linux Cluster의 MPICH에서 MPI_Allgather Algorithm의 성능을 개선하고 실험을 통해 최대 30%의 성능향상을 증명하였다. MPICH의 기존 버전이 메시지의 크기와 실행 프로세스 수에 따라 Recursive Doubling, Bruck Algorithm, Ring Algorithm을 차등 적용했던 것을, 앞의 Algorithm을 개선하여 Double Bruck Algorithm, Double Ring Algorithm을 제안, 구현하였다.

• Journal of KIISE:Computing Practices and Letters
• /
• v.9 no.6
• /
• pp.635-645
• /
• 2003

One of very important features that are necessarily supported by clustered parallel computer systems is a single I/O system image in which users can access both the local and remote I/O resources transparently. In this paper, we propose an efficient disk sharing technique supporting a single disk I/O system image architecture. The design separates the I/O subsystem of a cluster into the file system and a set of virtual hard disk drivers. The virtual hard disk driver deals with a hard disk in the remote node as a local hard disk. All services provided by it are performed in the device driver level without any modification of file systems. Users can, therefore, access all the disks in the cluster regardless of their locations. Our virtual hard disk driver is implemented under the linux, and also tested in a linux cluster system. We find by experiments that it can successfully support a single disk I/O space, and at the same time it shows better performance than NFS. We are sure that this paper can be a guideline for single I/O space of other devices to be easily constructed.

• The KIPS Transactions:PartA
• /
• v.11A no.5
• /
• pp.355-364
• /
• 2004

These days, various clustering technologies have been adopted to construct web sites. High performance hardware switches have good performance, but have disadvantage of high cost for constructing small and middle-sized web sites. Now a days, many sites have been constructed with the LVS (Linux Virtual Server), which is free of charge and has good performance. Having a centralized load balancing with one front-end, the LVS causes a bottleneck when it receives all at once. In the paper, we suggest a way to remove the LVS bottleneck by providing multiple front-ends. In this architecture, all of cluster nodes act as both a front-end and a back-end. When the load of a node receiving requests is not large enough, the node responds to the client directly. When the load of a node is large enough, the node send the request to a node which is selected by a scheduling algorithm. The scheduling algorithm is discussed to balance loads between servers. While single front-end cluster raises the throughput curvedly, the multiple front-end cluster raises the throughput linearly.

• Journal of computational fluids engineering
• /
• v.6 no.2
• /
• pp.47-55
• /
• 2001

In recent years cluster systems using off-the-shelf processors and networks components have been increasing popular. Since actual performance of a cluster system varies significantly for different architectures, representative in-house codes from major application fields were executed to evaluate the actual performance of systems with different combination of CPU, network, and network topology. As an example of practical CFD(Computational Fluid Dynamics) simulations, the flow past an Onera-M6 wing and the flow past an infinite wing were simulated on clusters of Linux and several other hardware environments.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2002.11c
• /
• pp.1769-1772
• /
• 2002

iBASE/Cluster는 한국전자통신연구원에서 개발한 객체 지향 멀티미디어 DBMS인 바다-IV를 기반으로 공유 디스크 기반의 클러스터 환경에서 무정지 DBMS 서비스를 지원하기 위하여 확장된 DBMS 이다. 본 논문에서는 이를 위하여 수행된 바다-IV 의 확장에 대해 기술한다. 기술된 내용은 공유 디스크 자원 접근을 동기화 하기 위한 동시성 제어 기법 및 버퍼 관리 기법, 그리고 특정 노드에 고장 발생 시 회복 기법 등에 대한 확장을 포함한다. 또한 확장된 DBMS가 LVS(Linux Virtual Server) 기반의 클러스터에서 자동적인 고장의 감지 및 이전(fail-over)을 수행하기 위한 연동 구조를 다루고 동작 시나리오를 통하여 구현된 시스템의 흐름을 보인다.

• 한국전산유체공학회:학술대회논문집
• /
• 2005.04a
• /
• pp.41-46
• /
• 2005

A new PC cluster was designed and constructed based on Windows XP Operating system. Primary target of the present design was the high node density per rack by using the general PC parts those are cost-effective and readily available in the market. Other major design points were system cooling and the convenient maintenance using standard PC parts. Presently 24 nodes per rack seems to be optimum considering the specification of the network switching device, system cooling and power supply, but 40 nodes can be accommodated within a single rack at maximum. Windows XP was selected as a high-performance computing environment considering the cost and the convenience in acquisition, maintenance and education. Both fast-Ethernet and Gigabit Ethernet network connection were tested and compared with previous data, especially for Linux doter using Myrinet. The result shows that there is no significant difference between the operating systems and the Fast-Ethernet and/or Gigabit Ethernet are good solution for the high-performance PC cluster considering the cost and performance.

• The KIPS Transactions:PartA
• /
• v.13A no.6 s.103
• /
• pp.517-524
• /
• 2006

In cluster computers, it is essential to Implement the single I/O space(SIOS) supporting integrated I/O substructure to efficiently process I/O intensive applications. SIOS service provides with global I/O address space to directly access peripherals and hard disks in its own or remote nodes from any node in the cluster computer In this thesis, we propose the implementation method of SIOS in Linux clusters by using only freewares. This method is implemented at device driver level that uses Enhanced Network Block Device(ENBD) and file system level that uses S/W RAID and NFS. The major strengths of this method are easiness of implementation and almost no cost due to using freewares. In addition, since freewares used are open sources, it is possible to apply this method to other platforms with only slight modification. Moreover, experiments show that I/O throughputs are up to 5.5 times higher in write operations and approximately 2.3 times higher in read operations than those of CDD method that uses the device driver developed at kernel level.

• Journal of the Korean Society for Aeronautical & Space Sciences
• /
• v.34 no.8
• /
• pp.33-40
• /
• 2006

A parallel computing strategy for finite element(FE) processing is described and implemented in nonlinear explicit FE code and its parallel performances are evaluated. A self-made linux-cluster supercomputer with 520 CPUs is used as a bench mark test bed. It is observed that speed-up is increased almost idealy even up to 256 CPUs for a large scale model. A communication over head and its effect on the parallel performance is also examined. Parallel performance is compare with the commercial code and developed code shows superior performance as the number of CPUs used are increased.