Copyright © Curt Hill, RAID What every server wants!

Copyright © Curt Hill, History I In the late part of the 1970s and there were two types of disks: –Hard –Floppy The hard drives were only used on mainframes or minis –Quite expensive The floppys were mostly used on personal computers In about 1984 hard drives started appearing on IBM personal computers

Copyright © Curt Hill, History II Now there were two types of hard drives: –Professional Large M Expensive –Amateur Small: 5-50 M Inexpensive

Copyright © Curt Hill, History III At this point economics takes charge The large disks were a small market, typically 10s of thousands The small disks were mass market, typically 10s of millions What happens to the prices?

Copyright © Curt Hill, History IV Someone gets a bright idea: –Should I buy one 500 M professional disk for $50,000 or 10 50M amateur disks at $100 each? RAID is born: Redundant Array Inexpensive Disks –Later Redundant Array Independent Disks There might even be a performance improvement if I can access the array independently However, we need some type of controller to treat the 10 disks as if it were one

Copyright © Curt Hill, Issues Redundancy –If the small is less reliable or if the data is so important we need multiple copies for safety Speed –Increase speed by writing part of the data to one drive and another part to another at the same time Versions –There have been many, termed Levels –Currently level 0 through level 6

Copyright © Curt Hill, Redundancy Multiple copies allows one disk to crash without losing data Two forms: –Mirroring AKA Shadowing, Duplexing –Error Correction Codes

Copyright © Curt Hill, Mirroring Write all the data twice –Once to a disk and its mirror –This is done simultaneously, so no extra delay A read only has to wait for the faster of the two If a disk crashes, no rebuild is needed since the other disk may just be copied High disk overhead, needs two disks to store one disk worth of data Can read different parts of the two at same time to increase speed

Copyright © Curt Hill, Error Correction Code Even more history and background is needed here See the ECC.ppt presentationECC.ppt Many of the EC codes were studied by Hamming of Bell Labs –Thus known as Hamming codes

Copyright © Curt Hill, ECC Instead of mirroring which requires double disk space use an ECC Conceptually: –The eight bit data placed on eight separate drives and the four bit ECC on another –Any one disk that fails may be recreated from the rest

Copyright © Curt Hill, Speed Speed requires parallelism –Do two things at once With a mirrored disk read the front half from one and the back half from the other Transfer time cut in half This leads to a more general approach: stripes

Copyright © Curt Hill, Stripes Cut the data into stripes If you have N disks –Partition into N pieces –Read or write N pieces at a time Best if each piece goes to a separate controller

Copyright © Curt Hill, Controllers Controllers may be hardware or software or both Some levels make a software controller a problem

Copyright © Curt Hill, RAID Levels Originally specified with six levels –0-5 Some of these are seldom used and new ones have been added

Copyright © Curt Hill, Level 0 Striped disk array Minimum of two disks No redundancy or error checking –If a drive is lost all the data is lost –Only RAID that does not protect data Real RAID or not?

Copyright © Curt Hill, Level 0 - Striping A1A2A3A4 B1B2B3B4 C1C2C3C4 Four stripes Each block of data is written in four pieces No redundancy

Copyright © Curt Hill, Level 1 Mirroring Minimum of two disks 100% redundancy –A rebuilt disk is just a copy, not computed Simple controller Cannot be expanded on the fly

Copyright © Curt Hill, Level 1 - Mirroring AAEE BBFF CCGG Four disks, each is a mirror of another Each block of data is written twice Need twice as much space

Copyright © Curt Hill, Level 2 Striped disk at bit level ECC provides the redundancy Minimum of seven disks for storing 4 bit word Not commercially viable Number of parity disks is proportional to log of number of data disks Not very flexible

Copyright © Curt Hill, Level 2 – Striping with ECC A1A2A3A4 B1B2B3B4 C1C2C3C4 Four stripes of data protected by three of ECC ECC is computed on the fly EAx EBx EAzEAy EByEBz ECxECyECz

Copyright © Curt Hill, Level 3 Striped disk array with one ECC disk Each block is striped across disks One disk may fail without diminishing throughput Minimum of three disks High data rates Controller should be in hardware, not just software

Copyright © Curt Hill, Level 3 – Striping with ECC A1A2A3A4 B1B2B3B4 C1C2C3C4 Four stripes of data protected by three of ECC ECC is computed on the fly ECCA ECCB ECCC

Copyright © Curt Hill, Level 4 Similar to level 3 except –Blocks are not subdivided –Different block are written to different disks Minimum of three disks Controller is complex –Should be hardware Not easy to rebuild in case of failure

Copyright © Curt Hill, Level 5 Striped disk array with distributed ECC blocks Each disk stores both data and ECC –A disk stores ECC data for other disks Minimum of three disks Any single disk failure will result in no loss of data Most complex controller design

Copyright © Curt Hill, Level 5 – Striping with Distributed ECC A1 A2 A3 A4 B1 B2 B3 B5 C1 C2 C5 C4 Four stripes of data Each ECC protects other four No disks are only data or only ECC ECC5 ECC4 ECC3 ECC2 ECC1D1 D5 D4 D3 E2 E3 E4 E5

Copyright © Curt Hill, Level 6 Striped disk array –Similar to five except two independent ECC schemes –ECC blocks distributed among data disks Minimum of four disks May have multiple disk failures without loss of data

Copyright © Curt Hill, Level 6 – Striping with Two Distributed ECCs A1 A2 A3 B1 B2 B3 C1 C3 C2 Four stripes of data Two types of ECC each of which protects a different group No disks are only data or only ECC ECC4 ECC3 ECC2 ECC1 D3 D2 D1 ECCb ECCc ECCdECCa

Copyright © Curt Hill, Others Several combinations of these exists –Mirroring a striped disk 10 –Striping a mirrored disk 53 –Level 0 and 3 –Level 0 whose stripes are level 3 arrays

What is coming? As disk arrays get larger the likelihood of a double failure becomes significant Although no product yet, it seems likely that a triple parity scheme is inevitable Copyright © Curt Hill,