초록 ▼

□ 사업의 목표 : 엑사스케일 초고성능컴퓨팅 구현 기술 확보 및 역량 강화
□ 주요사업 내용
○ 자체개발 초고성능 계산노드 안정화 및 대규모 계산자원 원격 제어 기술 개발
- 차세대 SSD기반 병렬 파일 시스템 설계를 위한 I/O 패턴 분석 및 워크로드 스케쥴링 기술 연구
- 초고성능 컴퓨팅 시스템 에너지 효율 분석을 위한 벤치마크 기술 연구
- HW 및 시스템 상태 모니터링 및 설정 제어 기술 개발
○ 대규모 시스템 에너지 효율성 제고를 위한 프로파일링 기술 개발
- 초병렬 고집적 컴퓨팅 시스

□ 사업의 목표 : 엑사스케일 초고성능컴퓨팅 구현 기술 확보 및 역량 강화
□ 주요사업 내용
○ 자체개발 초고성능 계산노드 안정화 및 대규모 계산자원 원격 제어 기술 개발
- 차세대 SSD기반 병렬 파일 시스템 설계를 위한 I/O 패턴 분석 및 워크로드 스케쥴링 기술 연구
- 초고성능 컴퓨팅 시스템 에너지 효율 분석을 위한 벤치마크 기술 연구
- HW 및 시스템 상태 모니터링 및 설정 제어 기술 개발
○ 대규모 시스템 에너지 효율성 제고를 위한 프로파일링 기술 개발
- 초병렬 고집적 컴퓨팅 시스템 최적화를 위한 프로파일링 도구 적용 및 고도화
- 초고성능 컴퓨팅 시스템 에너지 효율성 제고를 위한 데이터 수집 및 분석 기술 연구

(출처 : 초록 4p)

Abstract ▼

Ⅳ. Result of the study
○ Research for applying technologies of supercomputing
- Development of High Performance Computing Server Prototype
• Designing and making of high performance computing server based on latest Intel X86 Xeon processor
• Support 2-socket Cascade Lake-SP processor and

Ⅳ. Result of the study
○ Research for applying technologies of supercomputing
- Development of High Performance Computing Server Prototype
• Designing and making of high performance computing server based on latest Intel X86 Xeon processor
• Support 2-socket Cascade Lake-SP processor and 4 PCIe 3.0 slots for GPGPU accelerator
• All manufaturing process are including design, artwork, SMT, verification of main board and chassis design
• Board design, manufacturing, verification, and chassis manufacturing, etc. are carried out through cooperation withdomestic industries. Self-development ability improvement
• It is possible to improve stability and reliability in subsequent system development through SI analysis, cooling and flowanalysis analysis.
- Research on benchmark technology for energy efficiency analysis of high performance computing system
• Achieve a high level of energy efficiency with 5.93GFlops/watts (TTA Test and Certification Result Number: HPC-20-004)
- Development of control technology for monitoring and setting up the status of remote node of large-scale high performance system
• Web-based integrated management dashboard and management for efficient management
✔ Development of Web-based large-scale computational node remote management system
• Send and receive standard protocol-based messages for existing system compatibility
✔ IPMI message support for legacy server compatibility
• Standard message-based remote hardware monitoring and control
✔ Implementation of functionality of temperature, fan speed, power, FRU, BMC Watchdog
• Reliable hardware management, extended BMC capabilities, improved communication performance
✔ Scalable eXtension(SX) agent-based design and implementation for large scale remote management
- Study on I/O pattern analysis and workload scheduling for next-generation SSD-based parallel file system
• All-flash storage system performance analysis and improvement in HPC environment
✔ Evaluation of ZFS-based Lustre filesystem performance.
✔ I/O performance analysis according to stripe configuration and tunable parameter of ZFS.
✔ Comparative analysis of I/O performance of ZFS and Ldiskfs backend filesystem.
• Analysis on metadata performance and small file I/O performance of all-flash parallel file system
✔ Analysis on the effect of write amplification on metadata performance and small file I/O performance.
✔ Analysis on the effect of trim command via fstrim on WAF and write performance of NAND-flash SSD.
✔ Analysis on metadata performance and small file I/O performance with Data-on-MDT feature of Lustre.
✔ Proposal of MDT data separation scheme to reduce WAF of MDT.
• A study on memory I/O performance analysis of many core processors including high-performance on-chip memory
✔ I/O performance bottleneck analysis of MCDRAM of KNL processor

○ Development of system profiling analysis technology
- Application and advancement of profiling tools to optimize high parallel, high-intensity computing systems
• Development of advanced profiling technology for massively parallel and high-density computing system
✔ Build a profiling testbed cluster and determine the size(small/medium) of the application to be tested
✔ Performance analysis of performance event data collection in single node and multi-node clusters
✔ Derivation of application characteristics analysis results applying profiling techniques according to node size
✔ Analysis of scalability of profiling techniques through optimization of scheduling of large-scale cluster systems
• Advanced benchmark automation tool for high-performance computing cluster system
✔ Enhancement of dynamic benchmark automation tools with auto-scaling techniques and support for profiling data capabilities
✔ Comparison of various high-performance computing benchmark performance and power experiments based on KNLcluster
• Research on hardware performance counter data collection method and overhead in cluster system
✔ Development of a tool for collecting hardware performance counters of remote computing nodes in a cluster environment
✔ The overhead incurred when collecting hardware performance counters is almost independent of frequency and is highly correlated with the number of items collected
- Research on the improvement of resource utilization of cluster high performance computing system
• Research on energy and performance-related profiling data analysis technology for high performance computing systems
✔ Profiling data set configuration for energy efficiency of high performance computing system
✔ Research on generating system profiling data to reduce energy consumption
✔ Development of high performance computing system power prediction model and real-time power prediction technology
• Research on energy efficiency improvement technology for high performance computing system based on profiling data analysis
✔ Research on energy efficiency technologies for high performance computing systems that complement the shortcomings of existing power management technologies
✔ Energy efficiency algorithm-based power consumption and performance measurement, average execution time increased by 5% power consumption decreased by 34%
• Research on the method for improving energy efficiency of high performance cluster system
✔ Research on thread and data mapping policy based on profiling data
✔ Research on energy management policy applying DFS (Dynamic Frequency Scaling) to Uncore

(출처 : SUMMARY 14p)

목차 Contents

• 표지 ... 1제 출 문 ... 3보고서 초록 ... 4요 약 문 ... 5Summary ... 11Contents ... 19목차 ... 20표목차 ... 21그림목차 ... 22제 1 장 연구개발과제의 개요 ... 27 제 1 절 연구사업 목표 ... 27 제 2 절 연구사업 필요성 ... 27제 2 장 국내외 기술개발 현황 분석 ... 31 제 1 절 해외 동향 ... 31 제 2 절 국내 동향 ... 34제 3 장 연구개발 수행 내용 및 결과 ... 37 제 1 절 슈퍼컴퓨팅 활용목적형 구현 기술 확보 ... 37 제 2 절 시스템 프로파일링 분석 기술 개발 ... 72 제 3 절 지적 재산권 현황 ... 111제 4 장 결론 ... 113 제 1 절 대표 성과 ... 113 제 2 절 관련분야에의 기여도 ... 116 제 3 절 연구개발결과의 활용계획 ... 117참고문헌 ... 118부록 ... 123끝페이지 ... 223