HPC 자원최적화솔루션 (MOAB) + 사용자및관리자의생산성향상 Fall 2017 1
Adaptive Computing 대규모환경을지원하는 HPC job/workload 관리소프트웨어 Sources: IDC HPC End-user Study of System Software and Middleware in Technical Computing, Intersect360 Research HPC Site Census 스케쥴링및최적화부문의다양한혁신 대규모 HPC, Big Data 환경에적용되어있는 70+ 이상의지적재산권 Intel, HPE, Lenovo, Cray, SGI, ISBC 등다양한파트너보유 Fortune 500 및 Top 500 Systems 관리최적화 정유, 금융, 제조, 연구, 학교및정부기관 2
Moab Intelligence Engine + Torque Moab HPC Suite 15+ 년이상의대규모구성에서의검증 70+ 이상의지적재산권 현실에서의자원운용에대한의사결정지원 Multi-dimensional Policies Optimize Across: Application부하에대한요건 SLA와업무우선순위 시간 ( 실시간및미래예측 ) 이기종자원 Torque Open Source Resource Manager Moab HPC Suite 에통합운용 3
Intelligent Workload Optimization Moab HPC Suite Viewpoint or CLI Admin Dashboard APIs (Web Services, CLI, etc.) External System Report Interface Moab Intelligence Engine 의사결정, 정책, 스케쥴링, 배치, 및조직화 회계처리 Resource Managers Queue Mgr (Torque/Other) 모니터 배포 스토리지 네트워크 Non-traditional 부하 HW 자원 Other Resources: 라이선스 디스크 기타 4
Value Added Features / Applications HPC Suite 및추가가능한다양한 Add-on 모듈 유연한구성가능 HPC Suite Nitro High Throughput 5
핵심기능 6
생산성향상을위한 Viewpoint Portal 사용자 Faster and easier submission Reduce knowledge requirement Reduce errors Automate best practices into submission Users improve efficiency with system feedback 관리자 Manage users in less time Instant visibility into system and workload status Faster troubleshooting 7
MOAB View Point 8
통합된워크로드관리 이기종클러스터통합 독립된이기종자원을하나의클러스터로통합관리 하나의통합된시스템환경에서다중스텝작업을이기종클러스터연계하여실행가능 성능및사용률최대화 효율적으로통합된관리및정책운용가능 자원의가용성확보 Moab Grid Head Node All Rules for Grid Environment or Shared Grid Rules Moab Local Rules Moab Local Rules Moab Local Rules Frankfurt, Germany London, England New York, USA 9
유연한컴퓨팅환경 10
유연한컴퓨팅환경 지역공유자원추가삭제가능 기존클러스터에자원추가삭제 추가 : Backlog Completion Time 임계치 Example: If backlog > 600 seconds (10 minutes) add a specified number of nodes for a specified duration (a.k.a. request geometry) Option: Set request geometry to size & duration of highest priority job 삭제 : based on specified duration 혹은 Node Idle Purge Time Example: If node is idle > 300 seconds (5 minutes) remove node Administrators Moab Measure Remove Prepare Add Summary: Backlog: Nodes: Appear Wait Time as Extension: Improves: Request With Track Use Nodes Moab s pre-configured when nodes are the elastic released be backlog added computing images. when when the New Workload nodes appear runs and as reduces extension requested wait Use back queue triggers times log is time completion reduce greater for is post up & than or service install time nodes a of initial backlog. cluster. are exceeds specified levels configurations. idle > a improve. policy duration. threshold. limit. OpenStack Provisioning & Virtualization Workload Queue Moabconfigured VM Images HPC Users VM VM Requires: Externally supplied provisioning trigger setup. 11
Moab Accounting Manager 사용자, 그룹, 계정을위한사용예산총합설정 자원배치및스케줄링확인하여총합과비교 각각의사용자, 그룹, 계정에따른사용량에대한보고서및추적하여보여주거나환불처리진행 Set Usage Budgets $ 20% Engineering A Engineering Finance B $ Special Projects 15% $ 20% Analytics $ $ 20% 25% updates budget Checks budget Track usage & determine charges Budget & usage reports for showback or chargeback Job Request scheduler NODES SLA Allocates & schedules 12
Add On Nitro High Throughput Manager 대규모 HPC 를위한고성능 Job 배치확장 10 100배빠른 job 배치 (Up to 10/second/core) 신규 : 단일노드에서의 Multi-core serial 혹은 parallel jobs Workers Resource Moab Upon Nitro Summary: continuously allocates completion Start manager Up block script (Torque, of update all of Coordinator reads in the instantiates worker etc.) Nodes status Seamlessly High executes tasks, throughput via to Coordinator, the coordinator integrate the resource Nitro job on workload and assigns tasks first notifying terminates Start manager submitted high node Up throughput of and script (Torque, worker progress, worker (as instances etc.) it into Moab agents failure would for out to worker nodes on Nitro high remaining High and capability a job completion exits Throughput compute script) itself systems nodes job Moab Resource Manager Cluster N Zoomed View 13
Add On 고성능원격시각화 ( 원격워크스테이션 ) 생산성향상 Avoid waiting for file transfers Collaborate on shared projects Improve access to high value resources Local or Remote offices, home, etc. 자원저장 Avoid purchasing individual high end desktops Share high memory resources, accelerators, etc. Share expensive application licenses 관리통합 Easier to launch, find and manage sessions from submission portal Schedule, prioritize, and guarantee SLAs between users and groups Integrate into full HPC workflows, including notifications Administrative controls over usage, security, sharing, etc. 14
Admin Portal Unify Administration & Reporting 관리자보고서 워크로드상태모니터 단순화된 Job 자원뷰 15
Add On 보고서및분석 16
Add On 보고서및분석 17
MOAB 기본라이선스 18
ViewPoint 기본라이선스 19
20