초병렬 고집적 컴퓨팅을 위한 핵심 기술 개발

Abstract

□ 사업의 목표: 페타플롭스 성능의 슈퍼컴퓨터 독자 개발 구축과 슈퍼컴퓨터 시스템/운영 기술 확보를 통한 차세대 슈퍼컴퓨팅 소프트웨어 역량 강화

□ 주요사업 내용

○ 초병렬 고집적 컴퓨팅 시스템 기술 연구 개발 - 초병렬 고집적 컴퓨팅을 위한 시스템 원형 개발 - 고성능 인터커넥트 통신 기술 검증 및 개발 - 차세대 슈퍼컴퓨터 관리 운영 소프트웨어 기술연구 및 개발 ○ 매니코어 기반 슈퍼컴퓨터 작업 및 데이터 처리 기술 연구 - 매니코어 기반 작업 처리 기술 - 고성능 데이터 처리 기술 ○ 초고성능 컴퓨팅 활용 수치해석 라이브러리 개발 - 임의의 정방형 밀집행렬의 CSR 행렬변환 및 CSR 부분행렬의 선택적 밀집행렬 변환기능 수행 모듈 개발 - Conjugate Gradient 방법을 이용한 거대 CSR 행렬로 구성된 선형시스템 연산 모듈 개발

(출처 : 보고서 초록 3p)

Ⅳ. Result of the study ㅇResearch and Development of Massively parallel and high-density Computing Technology - Research and Development of massively parallel computing system using Intel's KNL processor · Designed a functional architecture of KNL processor main board and manufactured cluster system in collaboration with domestic hardware company. · Research on X86 platform firmware for KNL processor - Development and verification of high performance interconnect technology · Development of x8 and x16 PCIe adapter card based on on-board type optical module · Bandwidth performance test comparing with commercial PCIe adapter cards for effectiveness verification · Technical cooperation with Samtec to improve adapter card performance · Hardware environment construction and communication software technology development for Rack-Scale fabric interconnect switch · PCIe fabric switching chip based Top-of-Rack (ToR) switch prototype design and development for Rack-Scale interconnect · Device software and GPI development for PCIe interconnection network including development of symmetric memory technology · Verification of symmetric memory transmission technology through OpenSHMEM RDMA experiment · Development of performance improvement technique by applying Scatter-Gather DMA and implementation of PCIe-based OpenSHEMEM API subset - Next generation operation & management system Development · Research on system log analysis using machine learning technology · Improve stability and flexibility of management system · Architecture design for multi-cluster management · Monitoring and visualization of infiniband network · Design and deploy of container image repository to support docker ㅇResearch and Development of Many-core based Supercomputer Job and Data Processing Technology - Research on many-core processor based job processing technology · Development of parallel program profiling technology using next-generation many-core processor based performance counter · Development of classification technique based on characteristics of application using next-generation many-core processor profiling and data mining · Development of cluster-level job scheduling optimization technology based on analysis of application execution pattern for maximizing utilization of next-generation many-core processor - Research on high-performance I/O system · Analysis of the characteristic of I/O workload collected from large scale scientific application and Development of I/O log emulator in order to evaluate the hierarchical I/O system · Development of per-user I/O isolation in SSDs used in the burst buffer system · Construction of Lustre shared storage system based on LNET router in OmniPath-Infiniband mixed interconnect environment · Design of auto-provisioning technology of Lustre parallel filesystem through expansion of Ambari operational and management system ㅇ Library package for large-scale numerical analysis - Conversion of general dense matrices <--> CSR matrices · Procurement of the scalability and accuracy for the construction and store of large-scale matrices · Reduction of memory consumption needed to store arbitrary matrices - CG-based Large-scale linear system solver · Procurement of the scalability for calculations of large-scale linear systems · Strong performance in manycore systems in terms of speed and energy consumption, against the case of multicore systems · Independency on the type of manycore systems: Work in both PCI-E type devices (GPU) and onboard manycore systems (KNL).

(출처 : SUMMARY 13p)