국가 초고성능 컴퓨팅 리더쉽 제고를 위한 슈퍼컴퓨터 개발

Abstract

□ 사업의 목표: 페타플롭스 성능의 슈퍼컴퓨터 독자 개발 구축과 슈퍼컴퓨터 시스템/운영 기술 확보를 통한 차세대 슈퍼컴퓨팅 소프트웨어 역량 강화 □ 주요사업 내용 ○ 슈퍼컴퓨터 시스템 기술 연구 개발 - 슈퍼컴퓨터 자체 개발 프로토타입 설계 및 샘플 제작 - 고성능 인터커넥트 통신 기술 - 슈퍼컴퓨팅 관리 운영 소프트웨어 기술 ○ 매니코어 기반 슈퍼컴퓨터 작업 및 데이터 처리 기술 연구 - 매니코어 기반 작업 처리 기술 - 고성능 데이터 처리 기술 ○ 초고성능 컴퓨터 활용 수치해석 라이브러리 개발 - Compressed Sparse Row 알고리즘 기반의 표준 입출력 희소행렬 관리 모둘 개발 - Deflation LANCZOS 기반 고유값 계산 솔버 개발

Ⅳ. Result of the study ○ Research and Development of Supercomputer System Technology - Development of a prototype system for petaflops-scale supercomputer · Analysis of S7200AP(Adams Path) customer reference board(CRB) and functional requirements of computing node · Design and building of a prototype system for Intel KNL processor through cooperation with domestic companies specialized at hardware development · Research and verification for UEFI/BMC firmware · Study on OCP(Open Compute Project) systems and design of system rack cabinet

- Verification and development of high performance interconnection network technologies · Requirement analysis of PCIe-based switching chipset and optical module (i.e., optic component and active optic cable) to develop PCIe network adaptor card using Silicon Photonics · HW circuit design, PCB（printed circuit board) manufacturing, and performance/function verification of PCIe-based interconnection network adaptor using VCSEL on-board optic modules · Analysis of PCIe-based network switch and NTB (non-transparent bridge) technology for rack-scale interconnection network · Testbed development of PCIe NTB RDK (Rapid Development Toolkit) and proof-of-concept research of PCIe NTB-based communication software

- Next generation operation & management system Development · Research on Intel RSD based system monitoring and control system · Architecture design for cluster management · Survey and deploy OS provisioning technique and toolkits · Survey on dynamic software provisioning technology based on DevOps · Analysis of Open-source based computing/networking resource monitoring scheme · Research on system fault detection using machine learning technique · Implement of operation and management system for next generation supercomputer

○ Research and Development of Many-core based Supercomputer Job and Data Processing Technology - Research on many-core processor based job processing technology · Analysis on heterogeneous memory architecture of Intel Knights Landing (KNL) processor · Research on parallel application profiling technology by exploiting hardware performance counters · Over 7% performance improvement with no application code modifications by using multiple MPI application executions optimization strategy based on the profiling of parallel applications - Research on high-performance I/O system · Analysis and Applicability study of open-source software including GlusterFS, PLFS and OverlayFS for developing hierarchical I/O system · Designing a burst buffer system as an hierarchical I/O system · Building the burst buffer prototype and verify that the I/O throughput of new burst buffer is at least 200% greater than existing parallel file system and the performance is scalable according to the size of cluster through experiments on KISTI testbed.

○ Library package for large-scale numerical analysis - Sparse matrix storage based on the Compressed Sparse Row method · Management of non-zero elements in terms of number, start/end index of decomposed matrix blocks · Real-time update/delete of matrix elements · Dump of tentative results as CSR structures for a debugging purpose · Implemented as a MPI/C++ class - Deflation LANCZOS solver for computation of degenerated eigenvalues · Parallelization of (sparse)matrix-vector multiplier and vector dot-product · Modification of traditional Lanczos Iteration to handle multiple degenerated eigenvalues · Selective reorthogonalization of Lanczos-vectors to improve the accuracy of Lanczos Iteration · Performance improvement with offload computing : Simultaneous execution of Lanczos Iteration and T-matrix compouting in CPU and manycore devices, respectively.