Slide 1Slide 2Slide 3Slide 4Slide 5Slide 6Slide 7Slide 8Slide 9Slide 10Slide 11Slide 12Slide 13Slide 14Slide 15Slide 16Slide 17Slide 18Slide 19Slide 20Slide 21Slide 22Slide 23Slide 24Slide 25Slide 26Slide 27Slide 28Slide 29Slide 30Slide 31Slide 32Slide 33Slide 34Slide 35Slide 36Slide 37Slide 38Slide 39Slide 40Slide 41Slide 42Slide 43Slide 44Slide 45Slide 46Slide 47Slide 48Slide 49Slide 50Slide 51Slide 52Slide 53Slide 54Slide 55Slide 5615-410, F’04- 1 -Disk ArraysNov. 8, 2004Dave EckhardtDave EckhardtBruce MaggsBruce MaggsPresented by Michael Presented by Michael Ashley-RollmanAshley-RollmanL26_RAID15-410“...Failure is not an option...”15-410, F’04- 1 -SynchronizationToday: Disk ArraysToday: Disk ArraysText: 14.5 (a good start)Please read remainder of chapterwww.acnc.com 's “RAID.edu” pagesPittsburgh's own RAID vendor!www.uni-mainz.de/~neuffer/scsi/what_is_raid.htmlPapers (@ end)15-410, F’04- 1 -OverviewHistorical practicesHistorical practicesStriping, mirroringThe reliability problemThe reliability problemParity, ECC, why parity is enoughParity, ECC, why parity is enoughRAID “levels” (really: flavors)RAID “levels” (really: flavors)ApplicationsApplicationsPapersPapers15-410, F’04- 1 -StripingGoalGoalHigh-performance I/O for databases, supercomputers“People with more money than time”Problems with disksProblems with disksSeek timeRotational delayTransfer time15-410, F’04- 1 -Seek TimeTechnology issues evolve slowlyTechnology issues evolve slowlyWeight of disk headStiffness of disk armPositioning technologyHard to dramatically improve for niche customersHard to dramatically improve for niche customersSorry!Sorry!15-410, F’04- 1 -Rotational DelayHow fast How fast cancan we spin a disk? we spin a disk?Fancy motors, lots of power – spend more moneyProbably limited by data rateProbably limited by data rateSpin faster must process analog waveforms fasterAnalog digital via serious signal processingSpecial-purpose disks generally spin Special-purpose disks generally spin a littlea little faster faster1.5X, 2X – not 100X15-410, F’04- 1 -Transfer TimeTransfer time Transfer time Assume seek & rotation completeHow fast to transfer ____ kilobytes?How to transfer faster?How to transfer faster?15-410, F’04- 1 -Parallel Transfer?Reduce transfer time (without spinning faster)Reduce transfer time (without spinning faster)Read from multiple heads at same time?Read from multiple heads at same time?Practical problemPractical problemDisk needs N copies of analog digital hardwareExpensive, but we have some money to burnMarketing wants to know...Marketing wants to know...Do we have enough money to buy a new factory?Can't we use our existing product somehow?15-410, F’04- 1 -StripingGoalGoalHigh-performance I/O for databases, supercomputersSolution: parallelismSolution: parallelismGang multiple disks together15-410, F’04- 1 -StripingCa ffei en15-410, F’04- 1 -StripingStripe Stripe unitunit (what each disk gets) can vary (what each disk gets) can varyByteBitSector (typical)Stripe Stripe sizesize = stripe unit X #disks = stripe unit X #disksBehavior: “fat sectors”Behavior: “fat sectors”File system maps bulk data request N disk operationsEach disk reads/writes 1 sector15-410, F’04- 1 -Striping ExampleSimple case – stripe sectorsSimple case – stripe sectors4 disks, stripe unit = 512 bytesStripe size = 2KResultsResultsSeek time: 1X base case (ok)Transfer rate: 4X base case (great!)But there's a problem...But there's a problem...15-410, F’04- 1 -High-Performance StripingRotational delay Rotational delay gets worsegets worseStripe not done until fourth disk rotates to right placeI/O to 1 disk pays average rotational cost (50%)N disks converge on worst-case rotational cost (100%)Spindle synchronization!Spindle synchronization!Make sure N disks are always alignedSector 0 passes under each head at “same” timeResultResultCommodity disks with extra synchronization hardwareNot insanely expensive some supercomputer applications15-410, F’04- 1 -Less Esoteric Goal: CapacityUsers always want more disk spaceUsers always want more disk spaceEasy answerEasy answerBuild a larger disk!IBM 3380 (early 1980's)14-inch platter(s)Size of a refrigerator1-3 GByte (woo!)““Marketing on line 1”...Marketing on line 1”...These monster disks sure are expensive to build!Especially compared to those dinky 5¼-inch PC disks... Can't we hook small disks together like last time?15-410, F’04- 1 -Striping Example RevisitedSimple case – stripe sectors Simple case – stripe sectors 4 disks, stripe unit = 512 bytesStripe size = 2KResultsResultsSeek time: 1X base case (ok)Rotation time : 1X base case using special hardware (ok)Transfer rate: 4X base case (great!)Capacity: 4X base case (great!)NowNow what could go wrong? what could go wrong?15-410, F’04- 1 -The Reliability ProblemMTTF = Mean time to failureMTTF = Mean time to failureMTTF(array) = MTTF(disk) / #disksMTTF(array) = MTTF(disk) / #disksExample from original 1988 RAID paperExample from original 1988 RAID paperConner Peripherals CP3100 (100 megabytes!)MTTF = 30,000 hours = 3.4 yearsArray of 100 CP3100'sArray of 100 CP3100's10 Gigabytes (good)MTTF = 300 hours = 12.5 days (not so good)Reload file system from tape every 2 weeks???15-410, F’04- 1 -MirroringCopy ACopy B15-410, F’04- 1 -MirroringOperationOperationWrite: write to both mirrorsRead: read from either mirrorCost per byte Cost per byte doublesdoublesPerformancePerformanceWrites: a little slowerReads: maybe 2X fasterReliability Reliability vastlyvastly increased increased15-410, F’04- 1 -MirroringWhen a disk breaksWhen a disk breaksIdentify it to system administratorBeep, blink a lightSystem administrator provides blank diskCopy contents from surviving mirrorResultResultExpensive but safeBanks, hospitals, etc.Home PC users???15-410, F’04- 1 -Error CodingIf you are good at mathIf you are good at mathError Control Coding: Fundamentals & ApplicationsLin, Shu, & CostelloIf you are like meIf you are like meCommonsense Approach to the Theory of Error Correcting CodesArazi15-410, F’04- 1 -Error Coding In One Easy LessonData vs. messageData vs. messageData = what you want to conveyMessage = data
View Full Document