Is It Time For A New Storage Paradigm ? Network-Attached Storage
Transcript of Is It Time For A New Storage Paradigm ? Network-Attached Storage
Network-Attached Storage Devices:Network-Attached Storage Devices:Is It Time For A New Storage Paradigm ?Is It Time For A New Storage Paradigm ?
Paul Massigl iaPaul Massigl iaNetwork Storage Strategis tNetwork Storage Strategis t
Quantum Corporat ionQuantum Corporat ion500 McCarthy Boulevard500 McCarthy Boulevard
Milpitas, CA 95035Milpitas, CA 95035paulpaul.. [email protected]@quantum.com
7/24/99 Network Storage 99 2
TodayÕs I/O SystemsTodayÕs I/O Systems
nn A lot thatÕs good:A lot thatÕs good:ääcost-effective generic devicescost-effective generic devicesäämature, robust interconnects and softwaremature, robust interconnects and softwareääwell understood failure tolerance modelwell understood failure tolerance modeläämature device and file modelsmature device and file models
but. . .but. . .nn The model is not tracking technologyThe model is not tracking technologyääsystem topologiessystem topologiesääwire speed utilizationwire speed utilizationäädevice capabilitiesdevice capabilities DiskDisk
RAIDController
ClientClient Client
RAIDCon troller
ServerServer
7/24/99 Network Storage 99 3
Shortcoming 1:Shortcoming 1:Too Many ÒHopsÓ in the I/O PathToo Many ÒHopsÓ in the I/O Path
nn Each hop representsEach hop representsääa trip across a wireÉÉÉÉÉÉÉÉÉÉ.a trip across a wireÉÉÉÉÉÉÉÉÉÉ. fast and getting fasterfast and getting fasterääa Òstore and forwardÓ stepÉÉÉÉÉÉÉa Òstore and forwardÓ stepÉÉÉÉÉÉÉ minor slowdownminor slowdownääa Òsoftware stackÓÉÉÉÉÉÉÉÉÉÉ..a Òsoftware stackÓÉÉÉÉÉÉÉÉÉÉ.. bottleneckbottleneckçç wire protocol conversionwire protocol conversionçç operating system protocolsoperating system protocols
Server
RAID Controller
Client
Disk
Applica tionRequest
Ne tworkProt ocol
Ethernet
File Server(t ransformation)
RAID Cont roller(analysis)
FC
File Server(reduc tion)
ApplicationProcess
Ethernet
Ne tworkPro tocol
SCSI
RAID Con troller(consolidation)
FC
SCSI
t ime
"hops"
7/24/99 Network Storage 99 4
Shortcoming 2:Shortcoming 2:Not Well Suited To ClustersNot Well Suited To Clusters
nn Servers must be involvedServers must be involvedi ni näädata ownershipdata ownershipäädata access synchronizationdata access synchronizationääI/O load balancingI/O load balancingääfailure recoveryfailure recovery
nn This limitsThis limitsääscalingscalingääperformanceperformanceääavailabilityavailability
DiskDisk
How does oneRAID contr ollerknow what t heo ther is doing?
RAIDController
Who ownst he disks?
ClientClient Clien t
Who serveswhat data to
which clients?
RAIDController
What happenswhen something
breaks?
ServerServer
7/24/99 Network Storage 99 5
Enabler 1: Enabler 1: FibreFibre Channel Channel
nn PerformancePerformanceääbandwidth: 100 MBPSbandwidth: 100 MBPSäälatency: device-to-device data transferlatency: device-to-device data transfer
nn Availabil ityAvailabil ityäädual loop, ÒhotÓ plugging, zoned switchingdual loop, ÒhotÓ plugging, zoned switching
nn ScalingScalingäämore devices, longer distancesmore devices, longer distances
nn These have combined to create the ÒSANÓ conceptThese have combined to create the ÒSANÓ conceptnn Question: does the conventional Òblock serverÓ diskQuestion: does the conventional Òblock serverÓ disk
model utilizemodel utilize SANs SANs effectively? effectively?
7/24/99 Network Storage 99 6
Enabler 2: ÒSmarterÓ DevicesEnabler 2: ÒSmarterÓ Devices
nn Academic & industrialAcademic & industrialresearchresearchääSWARM --U of AZSWARM --U of AZääGFS--U of GFS--U of MinnMinnääPetal/Frangipani--DECPetal/Frangipani--DEC
SRCSRCääNASD/Active Disks CMUNASD/Active Disks CMUääetc.etc.
ääNASD/Active Disks CMUNASD/Active Disks CMU
7/24/99 Network Storage 99 7
Network Attached Secure DevicesNetwork Attached Secure Devices
nn PrinciplesPrinciplesäämake the disks ÒsmarterÓmake the disks ÒsmarterÓääwire clients directly to datawire clients directly to data
nn ChallengesChallengesääkeeping data securekeeping data secureääusing the ÒsmartÓ in the smart disks beneficiallyusing the ÒsmartÓ in the smart disks beneficially
ClientClientClient
"smart " disk "smart " disk "smart" disk "smart " disk "smart " disk "smar t" disk
robust network
7/24/99 Network Storage 99 8
A Precursor:A Precursor:MangoSoftÕsMangoSoftÕs MedleyMedley
nn ÒServer-lessÓ client-server computingÒServer-lessÓ client-server computingäädynamic disk resource sharingdynamic disk resource sharingääglobal file systemglobal file systemääredundancy--mirroring of data objectsredundancy--mirroring of data objects
nn Approximates the NASD modelApproximates the NASD modeläädisk intelligence is in the workgroup membersdisk intelligence is in the workgroup members
Local Space
Shared Space
Disk
Local Space
Shared Space
Disk
Local Space
Shared Space
DiskLocal Space
Shared Space
Disk
Network
7/24/99 Network Storage 99 9
File System FunctionsFile System FunctionsDisk Space ManagementDisk Space Management
Block 3724Block 3723
Block 3722
Block 0002Block 0001
Block 0000
Block 0121Block 0120
Block 0119
Block 1102Block 1101
Block 1100
Block 2598Block 2597
Block 2596
ll
l
Block 9999Block 9998
Block 9997
FredÕs mail (original)(extent 1)
other files and directories(not available for use)
free space(available for use)
Block 3723Block 3722
FredÕs mail (extended)(extent 2)
LBA space
ll
l
ll
l
ll
l
ll
l
7/24/99 Network Storage 99 10
File System FunctionsFile System FunctionsNamespace ManagementNamespace Management
DirectoriesData Direc tory
Fred’s Mail
Block 0022
Block 0002Block 0001
Block 0000
ll
l
Everyone’s Mail
FredÕs Mail
JanaÕs Mail
etc.
FileSystem
1997 Report
Block 0195
Block 0002Block 0001
Block 0000
ll
l
Corporate Data
1996 Report
1997 Report
etc.
System Stuf f
Security Log
etc.
Security Log
Block 0110
Block 0002Block 0001
Block 0000
ll
l
Applica tion’s view ofthe same disk
Block 9999Block 9998
Block 9997
Block 0002Block 0001
Block 0000
Block 2001Block 2000
Block 1999
l
l
l
l
l
l
File system’s view of a disk
7/24/99 Network Storage 99 11
File System FunctionsFile System FunctionsFile Access ControlFile Access Control
DiskDisk
Operating System
File System
no
yesDoes
user haveaccessrights?
Applicationrequest
foraccess to
a file
7/24/99 Network Storage 99 12
File System FunctionsFile System FunctionsData Access ControlData Access Control
DiskDisk
Operating System
File System
Applicationrequest
toaccess disk
directly
no
yesIs
requestedoperation
authorized?
Applicationrequest
toaccess data
in a file
no
yesDoes
user haveaccessrights?
Applicationrequest
foraccess to
a file
7/24/99 Network Storage 99 13
NASD: The Basic IdeaNASD: The Basic Idea
nn A NASD Òknows aboutÓA NASD Òknows aboutÓextentsextents
nn A ÒNASD File ManagerÓA ÒNASD File ManagerÓgrants a client access togrants a client access toan extentan extent
nn Data transfers directlyData transfers directlybetween NASD and clientbetween NASD and client
NASD
File ManagerDisablDisabl
eeExtentExtentAccessAccess
EnableEnableExtentExtentAccessAccess
Application (clien t )
CLOSECLOSEFILEFILE
O P E NO P E NFILEFILE
WRITEWRITEextentextentblocksblocks
R E A DR E A Dextentextentblocksblocks
I / OI / Ob u sb u s
I / OI / Ob u sb u s
7/24/99 Network Storage 99 14
NASD Functional PartitioningNASD Functional Partitioning
NASD NASD NASDNASD NASD NASDNASD
FileManager
ClientClientClient
robust network
nn Namespace managementNamespace managementnn File access controlFile access controlnn Data access controlData access control
nn Disk space managementDisk space management
7/24/99 Network Storage 99 15
The ÒSecureÓ in NASDThe ÒSecureÓ in NASD
nn File Manager and NASD share a secretFile Manager and NASD share a secretnn Secret is encrypted and handed to clie n tSecret is encrypted and handed to clie n tääKey point: File manager and NASD need not communicate toKey point: File manager and NASD need not communicate to
establish clientÕs access to a fileestablish clientÕs access to a file
NASD
Client
4. compareexpected &
received tokens
1. Request toaccess file x.dat
2. Capabiliy(file,access rights,
time, key)
3. request for data(capability)
5. Receive ordeliver data
7/24/99 Network Storage 99 16
NASD BenefitsNASD Benefits
nn Shorter I/O pat h sShorter I/O pat h säädata moves directly between storage and clientdata moves directly between storage and clientäädirect device-to-device transfers are possibledirect device-to-device transfers are possible
nn ÒSmartÓ devices should be able to self-optimizeÒSmartÓ devices should be able to self-optimizeääfor performancefor performanceääfor failure tolerancefor failure toleranceääfor securityfor security
nn No loss of functionNo loss of functionääall conventional file system functions can be implementedall conventional file system functions can be implemented
without negating NASD advantageswithout negating NASD advantages
7/24/99 Network Storage 99 17
Making NASD HappenMaking NASD Happen
nn Technical challengesTechnical challengesääfailure tolerancefailure toleranceçç serverserverçç networknetworkçç devicesdevices
ääsqueezing NASD functionality into disk-like costsqueezing NASD functionality into disk-like cost
nn Non-technical challengesNon-technical challengesääinstalled base transition from the block server modelinstalled base transition from the block server modeläämotivation for the incumbentsmotivation for the incumbentsçç RAID, OS, file system and database vendorsRAID, OS, file system and database vendors
7/24/99 Network Storage 99 18
Making NASD HappenMaking NASD Happen
nn There is progressThere is progress
nn NSIC (National Storage Industry Consortium) workin gNSIC (National Storage Industry Consortium) workin ggroupgroupääÒNSIC/NASDÓÒNSIC/NASDÓäämembers: CMU, Seagate, Quantum, IBM, members: CMU, Seagate, Quantum, IBM, StorageTekStorageTekäämeeting ~monthly to develop a NASD-like standards proposalmeeting ~monthly to develop a NASD-like standards proposalçç called ÒObject Based DisksÓcalled ÒObject Based DisksÓçç roughly: SCSI commands for NASD function with security deferredroughly: SCSI commands for NASD function with security deferred
ää~quarterly public meetings (next: Millbrae on 8/17/99)~quarterly public meetings (next: Millbrae on 8/17/99)äätarget: ANSI X3T10 proposal in fall 1999target: ANSI X3T10 proposal in fall 1999
7/24/99 Network Storage 99 19
WhatÕs ÒBeyondÓ NASDWhatÕs ÒBeyondÓ NASD
nn NASD isnÕt even ÒhereÓ yetÑwhy ask whatÕs beyond it?NASD isnÕt even ÒhereÓ yetÑwhy ask whatÕs beyond it?
nn NASD does two important thingsNASD does two important thingsääreplaces the conventional block access disk paradigmreplaces the conventional block access disk paradigmäädemonstrates the usefulness of intelligence in storage devicesdemonstrates the usefulness of intelligence in storage devices
nn but. . .but. . . a NASD is still a f ixed-function device a NASD is still a f ixed-function device
7/24/99 Network Storage 99 20
Beyond NASDBeyond NASD
nn What if storage device behavior could be adjusted in t h eWhat if storage device behavior could be adjusted in t h efield ?field ?
nn Would obviate questions of the ÒrightÓ devic eWould obviate questions of the ÒrightÓ devic ecapabil it ie scapabil it ie sääif a deviceÕs capabilities donÕt suit an application, downloadif a deviceÕs capabilities donÕt suit an application, download
different capabilitiesdifferent capabilities
nn Supporting researchSupporting researchääCMU: Active Disk ProjectCMU: Active Disk ProjectääUCB: ÒI-DiskÓ ProjectUCB: ÒI-DiskÓ ProjectääMicrosoft Research Group proposal: ÒMicrosoft Research Group proposal: ÒCyberbricksCyberbricksÓÓ
7/24/99 Network Storage 99 21
Sample Uses for ÒActiveÓ DisksSample Uses for ÒActiveÓ Disks
nn Data warehousingData warehousingää1000Õs of active disks search themselves simultaneously1000Õs of active disks search themselves simultaneously
nn DatabaseDatabaseää1000Õs of little transaction engines1000Õs of little transaction engines
nn HSM/backupHSM/backupääeach disk figures out what data needs to be moved or backed upeach disk figures out what data needs to be moved or backed up
nn Application customizationsApplication customizationsääÒsmartÓ caching, metered delivery, ÉÒsmartÓ caching, metered delivery, É
7/24/99 Network Storage 99 22
ConclusionConclusion
nn ItÕs time for the block access disk to evolveItÕs time for the block access disk to evolveääthe technology is therethe technology is thereääthe need is therethe need is there
nn CMU NASD work shows the possibil it iesCMU NASD work shows the possibil it iesääpotential benefits are greatpotential benefits are greatçç performance, scaling, system cost, upgrade granularityperformance, scaling, system cost, upgrade granularity
ääsignificant challenges remainsignificant challenges remain
nn The question:The question:ääwill NASD become the next ÒdiskÓ, or is a more radical storagewill NASD become the next ÒdiskÓ, or is a more radical storage
device paradigm change the right step?device paradigm change the right step?
7/24/99 Network Storage 99 23
NSIC/NASD and SNIANSIC/NASD and SNIA
nn NASD GoalsNASD Goalsäätimeliness: get a proposal to X3T10 in 1999timeliness: get a proposal to X3T10 in 1999ääqualityquality
nn Aligning with SNIA could be benefic ia lAligning with SNIA could be benefic ia läälots more perspectiveslots more perspectivesääpotentially lots more ÒcyclesÓpotentially lots more ÒcyclesÓ
nn Aligning with SNIA could be disrupt iveAligning with SNIA could be disrupt iveääbringing the new people up to speedbringing the new people up to speed
nn My ÒcharterÓ from NSIC/NASDMy ÒcharterÓ from NSIC/NASDääget us aligned without delaying the submissionget us aligned without delaying the submission