190 likes | 374 Views
Cloud Storage and Intel Power Management Usage Oriented Reference Architecture. Enrique Castro-Leon October 2009 Rev 1.0. Summary. The Cloud Computing Reference Architecture Model Historical context for Cloud Storage 2.0 Cloud storage yesterday, today and tomorrow
E N D
Cloud Storage and Intel Power Management Usage Oriented Reference Architecture Enrique Castro-LeonOctober 2009 Rev 1.0
Summary • The Cloud Computing Reference Architecture Model • Historical context for Cloud Storage 2.0 • Cloud storage yesterday, today and tomorrow • The growth of the digital universe and power considerations • Use cases for storage services producers and consumers • Usage Oriented Reference Architecture for Cloud Storage • CPU, servers, storage devices, rack, row, whole data center • How it works within and across architectural building blocks • Data center system-level use cases: platooning
Reference Architecture Model Usage: Discovery, Exploration phase Audience: C level Executives Content: High Level Conceptual Arch Example:Cloud Infrastructure RA General Purpose Reference Architecture (GPRA) High Level 30000’ Usage: Solution architecture discussion Audience: CTO, IT architects, IT Managers Content: Usage Specific/Environment Generic Example: Cloud Storage Power Management Usage Oriented Reference Architectures (UORA) Technical 1000’ Usage: Solution development/POC discussion Audience: IT architects & Engineers Content: Usage Specific/Environment Specific Example: EMC ATMOS Power Management Special Purpose Reference Architecture (SPRA) Mapping to Intel technology Deep technical End Users Usage: Strategic planning and product development Audience: Cloud Product Planning & Arch Forum Content: Solution Concept Proposals Example: Next Generation Power Management Solution Proposal Internal & FT ODC Platform Intercept Concepts Concept & Design TRP, PPOP, PLDM Usage: Defining Product Roadmap Audience:CPU and Platform Planners & Architects Content: Required Process Documentation Example: PPOP L2 Power Management Package Development & Implementation
Cloud Storage 2.0 **Source: John F. Gantzet al., “The Diverse and Exploding Digital Universe”, IDC March 2008
Cloud Storage Use Cases Cloud Storage Traditionaldedicatedcorporatestorage Storage Virtual InfrastructureServices In the Office Business UseCases – In-sourced vs. Out-sourced OnlineStorage Providers Corporatestorageservices Application &Data Storage Services RichInternetApplications Digital mediacreation &distribution DataPresenceService FamilyMovies-2-go Personalstore-n-sync SocialComputing Backup Streaming Other corpcloud-basedapps Storage-basedapplications Storage providers Storage consumers Consumer use cases (individual & corporate users
Challenges & Opportunities (consumers) • Challenges • Risk of data loss • Solve backup problem • Proliferation of storage media • Unwieldy data management • Data volume increasing • Digital media: audio, video, pictures, data • Media obsolescence • Data retrieval, migration • Opportunities • A la carte storage services • Data presence (Pi Corporation*), backup (Mozy*) • Data stewardship • DR, offsite storage
Challenges & Opportunities(corporate & service providers) • Challenges • Exponential growth of digital footprint • New media, Web 2.0 driving storage requirements • Cost of in-house data storage • Energy usage can’t be ignored • Opportunities • Virtualization driving storage usage/delivery • Service oriented IT with SLA-driven products • Retention, privacy, access, management, security • Service decoupled from underlying technology • Cloud based storage • Integrated power management for storage devices • Organization-wide, uniform policies • Advanced info capture, search, discovery, classification
Hierarchical approach to power management Power Control Range Hardware entity Implementation technology 100 KW Data Center OEM, ISV, SI solutions 10 KW Row Intel® Data Center Manager (10-10,000+ nodes) Rack 1 KW Intel® Dynamic Power Node Manager(single node) Power control through nested control loops Server 100 W Cache River (IPMI abstraction) Spinning/ SSD Disk Arrays Chipset CPU voltage & frequency scalingthrough control of ACPI P- and T-states Mem CPU 10 W
Cloud Storage Power Management Usage Framework Data Center Management Systems Rack Optimization Maximize server rack capacity in power constrained racks Branch CircuitOptimization Optimize DC pwrinfrastructReduce stranded power Platooning & workload rebalancing Optimize workload ramp-up, energy usage in virtualized environments Chargeback Fine grained reporting, billing, auditing P/T Monitoring Control power consumption against forecast targets Power control algorithms Power feedback loops & state estimation physical models Node level power management Management S/W (multi-level) Other P/T sensors, controls & algorithms ACPI P-state control Recursive composition Intel Confidential
Cloud Storage Systems Integration Framework Building Management System 1 Global policy: Appliance power not to exceed 5 KW Cloud storage application Cloud storage powerpolicy engine budgets storage and node power to meet 5KW objective Row Power Manager Rack PolicyEngine 3 Cloud storage policy engine manages storagepower throughselective spin-downs Storagecontrol Nodecontrol Cloud storage policy engine sets power targetfor compute & storagesubsystems Intel DCM technologycloud plug-in 5 Actual Powerconsumed Per node actual Powerconsumed 2 4 Storage subsystem set power 7 Compute subsystem set power DCM Disk ArrayEnclosure Reduce baselinepower in storage subsystem throughuse of SSDs Group Policies Node Manager performs voltage & frequency scaling to keep server power within assigned node power target PSMI PMBus PMBus PMBus Actual Powerconsumed Actual Powerconsumed Disk ArrayEnclosure NM IntelligentPDU S S CPU S Power tostorage subsystem DCM assigns power targets inindividual nodes dynamicallysubject to compute subsystem power constraints 6 Disk ArrayEnclosure
Cloud Storage Power Management Reference Architecture Data Center MgtBuilding Mgt System StorageApplication Presentation subsystem (consoles, dashboards, GUIs) Planning & Management CMDB Apache DCM Console REST SOAP DC Control … Event Mgr CMS Core Operational Control Native clientaccess API Policies Engine Resource monitoring & Mgmt Power Mgmt Discovery & provisioning Power control Policy Resolution Aggregator Policy Data Trending/ Query Data Node Adapter Service (plug-in framework) Monitoring & Automation Powercontrol Compute Subsystemset pwr &pwr readout(REST) Java in-process Info collection DCM Node control Intel® Data Center Manager Storage service Node Adapter Interface (pre-defined) Set power (IPMI) Power metadata (IPMI) Blade IB … 3rd-party NM DCMI iPDU iDRAC Set power & power metadata (REST) Set power & power metadata (REST) DCM Network Communication Lib CPU power control Powerreadout Tempsensor NPTM policyengine Monitors Power change Power PDU MetadataService Temp Mon PMBusP/S NM firmware Other environmental sensors (temp, airflow, humidity) PMBusinterface CPU power control PSU Mon BIOS ACPIpower & thermal state controlinterface Power Datastorage Metadatastorage Power control Actual powerconsumed Actual P-state Set P-state Operating systemor virtualization hypervisor P-state table KVM subsystem Cloud Storage Application Appliance Group Manager Node power control CPU
Cloud Storage Appliance-Level Replication Rack level power and environmental sensor data (SOAP) Cloudstorageapplication Cloudstorageapplication Cloudstorageapplication Cloudstorageapplication Rack PolicyEngine Rack PolicyEngine Rack PolicyEngine Rack PolicyEngine Storagecontrol Storagecontrol Storagecontrol Storagecontrol Nodecontrol Nodecontrol Nodecontrol Nodecontrol Row Power Manager Application powerpolicy (SOAP) Branch circuit and per appliance power allocation The Row Power Manager can control multiple appliances
Energy Management Use CasesPower vs. Energy Management Powersaved at time t Energy savedbetween t1 and t2is the area betweenthe two curves t t1 t2
Integrated Power Management • Power Mgt mechanisms have operational restrictions • Node Manager: high workload • DBS: low workload • NHM power proportional computing: Pbaseline limited to about 50% • Pooled virtualized environment • Combine mechanisms • Extend operational envelope • Lower Pbaseline • Example: Server platooning
Advanced Use Case: Server Platooning 2% - 25% Takes 1 to 5 minto bring node online Power consumption 70% 50% 100% Transitions forincreasing workload Transitions fordecreasing workload Active workloads Power capped Asleep (ACPI S3, S4 or S5) Active, high priorityworkloads Unconstrained power Idle Platooning state PMBus PSMI PMBus PSMI PMBus PMBus PMBus PMBus PMBus PMBus PMBus PSMI PSMI PMBus PMBus PMBus NM NM NM NM S S CPU S CPU S S S S S S S CPU CPU S S
Conclusions • Digital universe growing exponentially • Data spilling from the corporate data center into the cloud • Power usage growth slower but still unsustainable • Usage drives power management architecture • Paradigm of arbitrarily scalable systems • Composition & replication • Spans CPU to whole data center • Power control range from watts to hundreds of kilowatts • Power vs. energy management