UC Regents Fall 2004 © UCBCS 152 L18: Real Processor Walkthru I2004-11-02 Dave Patterson(www.cs.berkeley.edu/~patterson)John Lazzaro (www.cs.berkeley.edu/~lazzaro)CS 152 Computer Architecture and EngineeringLecture 18 – Real Processor Walkthru I www-inst.eecs.berkeley.edu/~cs152/1UC Regents Fall 2004 © UCBCS 152 L18: Real Processor Walkthru ILast Time: A Case for RAIDOriginal paper now on class website ...group When these two parmcs agree, the faded bu was a 0, othcrwtse it RAID levels 2,3, and 4 By stormg a whole transfer umt m a sector, reads was a 1 If the check drsk IS the fadure,Just read all the data drsks and store the group panty in the replacement drsk can be mdependent and operate at the maxrmum rate of a disk yet sull detect errors Thus the primary change between level 3 and 4 IS that WC Reducmg the check d&s toone per group (C=l) reduces the overhead cost to between 4% and 10% for the group stzes considered here The performance for the thud level RAID system is the same as the Level 2 RAID, but the effectrve performance per dtsk mcreases smce it needs fewer check d&s This reductron m total d&s also increases relrabdtty, but since It is shll larger than the useful hfehme of disks, this IS a minor pomt One advantage of a level 2 system over level 3 is that the extra check mformatton assocrated with each sector to correct soft errors IS not needed, mcreasmg the capactty per dtsk by perhaps 10% Level 2 also allows all soft errors to be corrected “on the fly” wnhout havmg to reread a sector Table IV summarizes the thud level RAID charactensncs and Figure 3 compares the sector layout and check d&s for levels 2 and 3 mterlcave data 4 Tran$er UIlllS a, b, c & d Level 4 Sector 0 &la Disk 1 MlTF Exceeds Useful Lrfenme Secwr 0 Data Disk 2 A a T 2 A Total Number of D&s owrhcad cost Useable Storage Capacity EventslSec Full RAID (vs Single Disk) LargeRecu& D/S Large Writes D/S Large R-M-W D/S Small Readr DISC Small Vyrites D/2sG Small R-M-W DISC G=lO (820,000 hrs or >90 years) 1 1OD 10% 91% EIficclency Per Disk L3 WIL2 WILl 91/S 127% 91% 91/S 121% 182% 91/S 127% 136% 09/S 127% 8% 05/S 127% 8% 09/S 127% 11% G=25 (346,000 hrs or 40 years) 104D 4% 96% Eflctency Per Disk w Lx2 WILI 96/S 112% 96% 96/S 112% 192% 96/S 112% 142% 041s 112% 3% 02/S 112% 3% 041s 112% 5% Table IV Characterrstrcs of a Level 3 RAID The L3lL2 column gives the % performance of L3 tn terms of L2 and the L3ILl column give; it in terms of LI (>loO% means L3 IS faster) The performance for the full systems IS the same m RAID levels 2 and 3, but since there are fewer check dtsks the performance per dnk tmproves Park and Balasubramaman proposed a thud level RAID system without suggestmg a partrcular applicauon park861 Our calculattons suggest tt 1s a much better match to supercomputer apphcatrons than to transacuon processing systems This year two disk manufacturers have announced level 3 RAIDS for such apphcanons usmg synchronized 5 25 mch disks with G=4 and C=l one from IvIaxtor and one from Mtcropohs [Magmms 871 This thud level has brought the rehabrhty overhead cost to its lowest level, so in the last two levels we Improve performance of small accesses w&out changmg cost or rehabrlny 10. Fourth Level RAID Independent ReadsbVrltes Spreadmg a transfer across all &sks wuhm the group has the followmg advantage . Large or grouped transfer ttme IS reduced because transfer bandwulth of the entue array can be exploned But it has the followmg drsadvantagek as well . ReadmgAvnhng to a disk m a group requues readmg/wnhng to all the d&s m a group, levels 2 and 3 RAIDS can perform only one I/O at a Pme per group . If the disks are not synchromzed, you do not see average seek and rotattonal delays, the observed delays should move towards the worst case, hence the S factor m the equatrons above This fourth level RAID improves performance of small transfers through parallehsm--the abrhty to do more than one I/O per group at a ume We no longer spread the mdtvtdual transfer informanon across several &sks, but keep each mdrvrdual unit ma smgle disk The vutue of bit-mterleavmg 1s the easy calculatron of the Hammmg code needed to detect or correct errors in level 2 But recall that m the thud level RAID we rely on the drsk controller to detect errors wnhm a single drsk sector Hence, rf we store an mdrvrdual transfer umt in a single sector, we can detect errors on an mdtvtdual read without accessing any other drsk Frgure 3 shows the different ways the mformatron is stored in a sector for Sectar 0 L&a Dtsk 3 Sector 0 D& Disk 4 Sector 0 Check Disk 5 Sector 0 Check Ask 6 Sector 0 Check Disk 7 aEcc0 bECC0 CECCO dECC0 aEcc1 bECC1 cECC1 dEcc1 aEcc2 bECC2 cECC2 dECC2 ECCa ECCb ECCc ECCd (Only one check &Sk tn level 3 Check rnfo ts calculated aver each lran.$er 10~1 (Each @tier umt 1s placed tnto a single sector Note that the check ~$0 IS now calculated over a pece of each tran$er urut ) D I s L Frgure 3 Comparrson of locatton of data and check mformatlon In sectors for RAID levels 2, 3, and 4 for G=4 Not shown IS the small amount of check mformatton per sector added by the disk controller to detect and correct soft errors wlthm a sector Remember that we use physical sector numbers and hardware control to explain these ideas but RAID can be unplemented by sofmare ucmg logical sectors and disks At fust thought you mrght expect that an mdrvldual wnte to a smglz sector stdl mvolves all the disks m a group smce (1) the check disk mutt be rewritten wnh the new panty data, and (2) the rest of the data dash> must be read to be able to calculate the new panty data Recall that each panty bit IS Just a smgle exclusive OR of s+l the correspondmg data NIL 11 a group In level 4 RAID, unhke level 3, the panty calculatron is ITXFI simpler since, if we know the old data value and the old parity balue al well as the new data value, we can calculate the new panty mforrmror: sr follows new panty = (old data xor new data ) xor old pantv In level 4 a small wnte then uses 2 dtsks to perform 4 accesses-2 rea& and 2 wrnes--whtle a small mad mvolves only one read on one disk Table V summarmes the fourth level RAID charactensucs Note that all small accesses improve--dramatrcally for the reads--but the small read-modrfy-wnte is strll so slow relatrve to a level 1 RAID that ns applrcabduy to
View Full Document
Unlocking...