• Journal of KIISE:Computer Systems and Theory
• /
• v.27 no.5
• /
• pp.496-505
• /
• 2000

Clusters of workstations (COW) are becoming an attractive option for parallel scientific computing, a field formerly reserved to the MPPs, because their cost-performance ratio is usuallybetter than that of comparable MPPS, and their hardware and software can be easily enhanced to thelatest generations. In this paper we present the design and implementation of our runtime library forclusters of workstations, called "Collective I/O Clustering". The library provides a friendlyprogramming model for the I/O of irregular applications on clusters of workstations, being completelyintegrated with the underlying communication and I/O system. In the collective I/O clustering, two I/Oconfigurations are possible. In the first I/O configuration, all processors allocated can act as I/Oservers as well as compute nodes. In the second I/O configuration, only a subset of processors canact as I/O servers, The compression and software caching facilities have been incorporated into thecollective 1/0 clustering to optimize the communication and I/O costs. All the performance results wereobtained on the IBM-SP machine, located at Argonne National Labs.

• International journal of advanced smart convergence
• /
• v.11 no.2
• /
• pp.1-6
• /
• 2022

With the emergence of the new era of the 4th industrial revolution, compute-bound workloads with large memory footprint like big data processing increase dramatically. Even in such compute-bound workloads, however, we observe bulky I/Os while loading big data from storage to memory. Although file I/O cache plays a role of accelerating the performance of storage I/O, we found out that the cache hit rate in such environments is not improved even though we increase the file I/O cache capacity because of some special I/O references generated by compute-bound workloads. To cope with this situation, we propose a new file I/O cache management policy that improves the cache hit rate for compute-bound workloads significantly. Trace-driven simulations by replaying file I/O reference logs of compute-bound workloads show that the proposed cache management policy improves the cache hit rate compared to the well-acknowledged CLOCK algorithm by a large margin.

• Journal of the Institute of Electronics and Information Engineers
• /
• v.52 no.9
• /
• pp.54-62
• /
• 2015

Although NAND flash-based SSD (Solid-State Drive) provides superior performance in comparison to HDD (Hard Disk Drive), it has a major drawback in write endurance. As a result, the lifetime of SSD is determined by the workload and thus it becomes a big challenge in current technology trend of such as the shifting from SLC (Single Level Cell) to MLC (Multi Level cell) and even TLC (Triple Level Cell). Most previous studies have dealt with wear-leveling or improving SSD lifetime regarding hardware architecture. In this paper, we propose the optimal configuration of host I/O stack focusing on file system, I/O scheduler, and link power management using JEDEC enterprise workloads in terms of WAF (Write Amplification Factor) which represents the efficiency perspective of SSD life time especially for host write processing into flash memory. Experimental analysis shows that the optimum configuration of I/O stack for the perspective of SSD lifetime is MinPower-Dead-XFS which prolongs the lifetime of SSD approximately 2.6 times in comparison with MaxPower-Cfq-Ext4, the best performance combination. Though the performance was reduced by 13%, this contributions demonstrates a considerable aspect of SSD lifetime in relation to I/O stack optimization.

• Proceedings of the Korean Information Science Society Conference
• /
• 2011.06b
• /
• pp.423-426
• /
• 2011

SSD는 HDD와 다른 물리적 특성으로 인해 새로운 I/O 스케줄러를 요구하고 있다. 본 논문은 기존 CFQ 스케줄러가 SSD에서 어떤 성능을 보이는지 다양한 설정값 하에서 실험하고, 그 결과를 분석하여 SSD에서의 I/O 스케줄링 성능의 특징을 알아보고자 한다. 대상 설정값으로는 역탐색 제한 설정, FIFO 큐 내의 I/O 요청에 대한 만료 시간, I/O 요청 큐 별 할당 시간, 동기적 요청 큐를 위한 대기 시간을 고려하였다. 이러한 설정값을 다양하게 변화시키며 여러 개의 I/O 벤치마크 프로세스를 동시에 수행한 결과, 역탐색 제한을 제거하면 처리량이 상승하지만 그 폭이 매우 적으며, I/O 요청 만료 시간에 대해서는 연관성을 찾지 못하였다. I/O 요청 큐 별 할당 시간과 동기적 요청 큐를 위한 대기 시간은 클수록 처리량이 상승함을 보였다. CFQ의 목표인 평등성은 항상 97% 이상 보장됨을 관찰할 수 있었다. 우리는 이러한 결과가 차후 평등성을 보장하는 새로운 SSD I/O 스케줄러의 연구를 위한 초석이 되리라 기대한다.

• Journal of KIISE
• /
• v.43 no.7
• /
• pp.736-743
• /
• 2016

Several studies have reported the impact of core affinity on the network I/O performance of multi-core systems. As the network bandwidth increases significantly, it becomes more important to determine the effective core affinity. Although a framework for dynamic core affinity that considers both network and disk I/O has been suggested, the multiple queues provided by high-speed I/O devices are not properly supported. In this paper, we extend the existing framework of dynamic core affinity to efficiently support the multiple queues of high-speed I/O devices, such as 40 Gigabit Ethernet and NVM Express. Our experimental results show that the extended framework can improve the HDFS file upload throughput by up to 32%, and can provide improved scalability in terms of the number of cores. In addition, we analyze the impact of the assignment policy of multiple I/O queues across a number of cores.

• Journal of KIISE
• /
• v.42 no.2
• /
• pp.169-176
• /
• 2015

The atomic write technique is a good solution to solve the problem of the double write buffer. The atomic write technique needs modified I/O subsystems (i.e., file system and I/O schedulers) and a special SSD that guarantees the atomicity of the write request. In this paper, we propose the writing unit aligned block allocation technique (for EXT4 file system) and the merge prevention of requests technique for the CFQ scheduler. We also propose an atomic write-supporting SSD which stores the atomicity information in the spare area of the flash memory page. We evaluate the performance of the proposed atomic write scheme in MariaDB using the tpcc-mysql and SysBench benchmarks. The experimental results show that the proposed atomic write technique shows a performance improvement of 1.4~1.5 times compared to the double write buffer technique.

• Proceedings of the KIEE Conference
• /
• 2002.07a
• /
• pp.265-267
• /
• 2002

This paper deals with an I/O simulator design for performance evaluation of reactor protection systems in nuclear power plants. The I/O simulator provides input signals for the reactor protection system, and acquires output signals from the initiation circuits. The simulator is based on VMEbus system, and all VMEbus boards are developed within the country.

• Proceedings of the Korean Information Science Society Conference
• /
• 2010.06b
• /
• pp.460-464
• /
• 2010

SSD는 뛰어난 성능으로 인해 서버 시장에서 HDD를 빠르게 대체하며 각광받고 있다. 우리는 기존 SSD의 성능 분석이 단일한 I/O 패턴에 대해서만 이루어진 점을 주목하여, 다양한 패턴의 I/O가 동시에 수행 될 경우, 성능에 어떠한 영향이 있는지 평가해보고자 한다. 이를 위해 4KB부터 64MB까지 다양한 블록크기로 순차적/임의적 읽기/쓰기 연산을 수행함과 동시에 4KB 단위의 읽기/쓰기 I/O를 수행시켜 성능에 미치는 영향을 알아보았다. 이러한 평가를 네 가지 리눅스 I/O 스케줄러에 대해 각각 수행함으로써 스케줄러에 의한 영향 또한 평가하였다. 그 결과로 우리는 새로운 SSD의 성능 특성을 발견할 수 있었으며, 이는 새로운 I/O 스케줄러 및 SSD의 FTL 개발의 기반이 되리라 예상된다.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2010.11a
• /
• pp.1774-1776
• /
• 2010

Solid State Device(SSD)는 플래시 메모리 모듈을 기반으로한 차세대 저장장치로서 디스크 회전방식의 Hard Disk Drive(HDD)를 대체하는 장치로 주목받고 있다. 하지만, 현재의 운영체제 I/O 스케줄러는 HDD 에 최적화되어 있기 때문에 플래시 모듈 기반의 SSD 의 성능을 최대한 끌어내지 못하는 한계를 갖고 있다. 따라서, 본 논문에서 우리는 SSD 의 성능을 끌어올리기 위해 SSD 의 특징을 분석하고 이를 바탕으로 SSD 에 최적화된 I/O 스케줄러를 고안하였다. 새로운 스케줄러(NSCHED)는 I/O Request 를 두 그룹으로 분류하고 각 요청에 타이머를 설정함으로써 I/O Request 를 빠르게 처리함과 동시에 기아방지를 위한 기법을 가진다. 우리는 Linux 2.6.30 에서 NSCHED 스케줄러를 구현하여 Postmark 벤치마크를 통해 성능평가를 수행했으며, NSCHED 스케줄러가 기본 Linux I/O 스케줄러보다 30% 가량 낮은 응답 시간이 소요됨을 확인했다.

• 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.