초록 ▼

□ 사업의 목표 : 엑사스케일 초고성능컴퓨팅 구현 기술 확보 및 역량 강화

□ 주요사업 내용
○ 슈퍼컴퓨팅 활용 목적형 구현 기술 확보
- 자체 설계한 HDCA(High Density Compute Accelerator) 형태의 확장박스 구성의 자체개발한 UEFI BIOS, BMC 기술을 탑재한 초고성능 서버 6 노드 추가 제작 및 안정화
- 무스위치 PCIe 네트워크 환경에서의 OpenSHMEM 라이브러리 개발
○ 시스템 프로파일링 분석 기술 개발
- 다양한 벤치마크 도구를 하나의 성능 벤치

□ 사업의 목표 : 엑사스케일 초고성능컴퓨팅 구현 기술 확보 및 역량 강화

□ 주요사업 내용
○ 슈퍼컴퓨팅 활용 목적형 구현 기술 확보
- 자체 설계한 HDCA(High Density Compute Accelerator) 형태의 확장박스 구성의 자체개발한 UEFI BIOS, BMC 기술을 탑재한 초고성능 서버 6 노드 추가 제작 및 안정화
- 무스위치 PCIe 네트워크 환경에서의 OpenSHMEM 라이브러리 개발
○ 시스템 프로파일링 분석 기술 개발
- 다양한 벤치마크 도구를 하나의 성능 벤치마크 프로그램으로 통합하여 설치, 환경 설정, 구동까지 Non-Stop 실행 도구 개발
- 루프라인 모델 기반 HBM(High-Bandwidth Memory) 최적 활용 기술 및 HBM 동적 이주 기술 개발

(출처 : 초록 4p)

Abstract ▼

Ⅳ. Result of the study
○ Performance Improvement Study for PCIe Interconnect
- Research on DMA optimization of PCIe interconnetion network
• Find out optimization method to enhance the DMA performance of PCIe switch chip and improve the bandwidthby 160%
• Achieve 9.1 GB/s bandwidth by us

Ⅳ. Result of the study
○ Performance Improvement Study for PCIe Interconnect
- Research on DMA optimization of PCIe interconnetion network
• Find out optimization method to enhance the DMA performance of PCIe switch chip and improve the bandwidthby 160%
• Achieve 9.1 GB/s bandwidth by usgin multiple DMA engines simultaneously
- Performance evaluation of transmission cables for PCIe expansion
• Conduct performance evaluation for two types of PCIe expansion cables; copper type and optic type
- Study on improvement of scalability and stability of NTB based switchless network
• Improve the scalability and stability of switchless networks using PCIe switching-based optical adapter cards
• Enhance the RDMA throughput for PCIe network system based on PCIe 3.0 x8 interface
• Improved performance of non-transparent bridge (NTB) based software libraries for switchless networks
• Obtain scalable OpenSHMEM library software that works reliably on NTB-based switchless networks

○ Research on I/O Performance and Bottleneck Analysis in Manycore CPU System
- I/O performance and bottleneck analysis on KNL server
• Comparison of Buffered/Direct I/O performance between KNL server and Xeon server
• Performance Analysis of Buffered I / O of KNL According to MCDRAM Mode
• Analysis on the bottleneck caused by memory cgroup management function with Linux kernel I/O path profiling
• I/O performance analysis caused by different branch prediction accuracy between KNL and Xeon servers
• Performance Scalability Analysis based on KNL Cluster Mode in Multi-threaded I/O situation
- I/O performance analysis on EPYC server
• Buffered and direct I/O performance analysis in multi/single-threaded environment
• Analysis the effect of NUMA control setting on I/O performance
- Research for exploiting multi-streamed SSD on KNL servers
• Development of a kernel patch for utilizing multi-streamed SSD on KNL nodes based on RHEL 7 OS
• Analysis of the performance stability of multi-streamed SSD in KNL Server

○ Development of system profiling analysis technology
- Development of profiling technology for performance optimization of high performance computing systems
• Research on effective methods to apply hardware performance counter based system profiling data to improve performanceefficiency
✔ Application profiling and characterization in various systems using hardware performance counters
✔ Performance enhancements of systems through the resource interference-aware simultaneous scheduling algorithm between applications (29% on Intel Xeon-Phi system, and 15% on IntelXeon-sp system)
• Research on sampling method and power management technology for improving cluster system resource utilization
✔ Optimal Dynamic Voltage and Frequency Scaling (DVFS) Selection techniques for target power using profiling data(Run time 1.71% increase, power up to 42% reduction)
• Development of profiling tool applying analysis method suitable for system profiling data characteristics
✔ Development of system profiling tool prototypes that automatically run applications on high-performance cluster systems and automatically collect hardware performance counters and monitoring information
✔ Research on the lightweighting of cluster computing system massive monitoring data collection
✔ Creating system analysis reports using a self-developed system profiling tool
• Research on a data migration method for optimizing hybrid memory system
✔ Selecting data to be migrated through the loopline model based on monitoring information of running processes, andperform data migration between hybrid memories
✔ Applicable to various hybrid memory systems without changing source code
- Research on benchmark technology and profiling data collection technology for performance measurement of cluster environment
• A Study on Sampling Method and Power Management Technology to Improve Cluster System Resource Utilization
• Research and Development of Benchmark Technology for High-Performance Cluster System
• A Study on Profiling Data Collection and Analysis Method for High-Performance Computing Cluster System
• Research on clustering high performance computing system profiling data collection and analysis method
✔ Development of a profiling data-based performance analysis model
✔ Profiling data sampling rate based on raw data through profiling data sampling model of dimension reductionmethod 58.44%
✔ Performance counter-based profiling data up to 99% increase in accuracy of classifying learning data when generating training data, up to 14% increase in sampling data accuracy

(출처 : SUMMARY 14p)

목차 Contents

• 표지 ... 1
• 제 출 문 ... 3
• 보고서 초록 ... 4
• 요 약 문 ... 5
• Summary ... 11
• Contents ... 18
• 목차 ... 19
• 표목차 ... 20
• 그림목차 ... 22
• 제 1 장 연구개발과제의 개요 ... 27
• 제 1 절 연구사업 목표 ... 27
• 제 2 절 연구사업 필요성 ... 27
• 제 2 장 국내외 기술개발 현황 분석 ... 31
• 제 1 절 해외 동향 ... 31
• 제 2 절 국내 동향 ... 34
• 제 3 장 연구개발 수행 내용 및 결과 ... 36
• 제 1 절 슈퍼컴퓨팅 활용목적형 구현 기술 확보 ... 36
• 제 2 절 시스템 프로파일링 분석 기술 개발 ... 73
• 제 3 절 지적 재산권 현황 ... 139
• 제 4 장 결론 ... 141
• 제 1 절 대표 성과 ... 141
• 제 2 절 관련분야에의 기여도 ... 143
• 제 3 절 연구개발결과의 활용계획 ... 144
• 참고문헌 ... 145
• 부 록 ... 152
• 끝페이지 ... 265