Parity Logging O vercoming the Small Write Problem in Redundant Disk Arrays

1. Parity LoggingOvercoming the Small Write Problem in Redundant Disk Arrays Daniel Stodolsky Garth Gibson Mark Holland

2. Contents • Overview of some Raid systems • Small write problem • Parity logging • Floating data and parity • Comparison between different models • Concluding remarks • Questions

3. RAID systems consideredin this paper.

4. Small Write Problem • RAID 5 Small write may require prereading old data, writing new data, prereading corresponding old parity value, and writing new parity value. • RAID level 5 ,therefore, is penalized by a factor of four over nonredundant arrays for workloads of mostly small writes. • Mirrored disks are only penalized by a factor of two since data only needs to be written to two separate disks

5. OLTPand Small write • OLPT (On-line transaction processing) systems represent a substantial segment in of the secondary storage market . Bank System is an example • OLTP systems require update-intensive database services • Performance of OLTP is largely determined by small write performance.

6. Disk Bandwidth • The three components of disk access are: seek time, rotational positioning time, and data transfer time. • Small disk writes make inefficient use of disk bandwidth • Random cylinder accesses move data twice as fast as random track accesses which, in turn, move data ten times faster than random block accesses.

7. Parity Logging • A powerful mechanism for eliminating small write penalty. • Based on the much higher disk bandwidth of large accesses over small • A technique for logging or journaling events to transform small random accesses into large sequential accesses to log and parity disks

8. Basic Parity Logging Model • A RAID level 4 disk array with one additional disk, a log disk. • parity update image is held in a fault tolerant buffer • When enough parity update images are buffered, they are written to the end of the log on the log disk. • When the log disk fills up, the out-of-date parity and the log of parity update information are read into memory. • The out-of-date parity is updated (in memory) and rewritten with large sequential writes.

9. Basic Parity Logging Model

10. Reliability of Basic Logging Model • Data disk failure => • update parity disk • Reconstruct the lost data • Log or Parity disk failure • Install new empty log disk (or parity disk) • Reconstruct parity

11. Tracks, Cylinders, and Sectors

12. Parity Maintenance Time analysis (basic model vs Raid 4) • Every D small writes issued cause one track write to the log to occur • Every TVD small writes issued cause the log disk to fill up then 3 full disk accesses at cylinder data rate • => parity writes for TVD small writes consumes as much disk time as TV(D/10) + 3V(T/2xD/10) = TVD/4 • Result “Parity consumed by the parity update I/Os is reduced by about a factor of eight

13. Enhancing Basic parity Logging Model • Limitation • The Basic Parity Logging model is completely impractical since an entire disk’s capacity of random access memory is required to hold the parity during the application of the parity updates. • Enhancement (Parity Logging Regions) • dividing the array into regions. • Every region is treated the same way as an entire disk in the basic model • Each region has its own fault tolerant buffer

14. Parity Logging Regions

15. Enhancing Parity Logging Regions • Limitation • Log and parity disks may become performance bottlenecks if there are many disks in the array. • Enhancement (Log and parity Rotation) • Distributing parity and Logs across all the disks in the array

16. Log and parity Rotation

17. Enhancing Log and parity Rotation • Limitation • The log and parity bandwidth for a particular region is still that of a single disk. • Enhancement (Block Parity Striping) • Distributing the parity log for each region over multiple disks.

18. Block Parity Striping

19. Analytical Model • Single small write access in parity logging will on average take Which can be simplified to S + (3 + 2/D) R Without preread S + (1 + 2/D) R • More analysis • Writing fault tolerant buffers to Parity log regions. • Log parity integration

20. Simulation Parameters

21. Parity Logging Overheads vs RAID 5 Overhead (per small write) • Contributions to disk busy time for the example disk array ( previous slide) • Extra I/O done by RAID 5 cost nearly 35 milliseconds

22. Alternative SchemesFloating Data and Parity • Organizing data and parity into cylinders that contain either data only or parity only and • Maintaining a single track of empty space per cylinder

23. Floating Data Parity

24. Floating Data and Parity (analysis) • For RAID 5, busy time for each data and parity update is S + R + 2R/D + (2R – 2R/D) + 2R/D • With new technique (2R – 2R/D) term is replaced with a head switch and a short rotational delay ( 0.76 data units using the sample array mentioned before) • Small random write in floating data and parity is 2S+(2+11.04/D)R + 2H • This is close to mirroring performance if D is large and H is small

25. Model Estimates (as predicted by analysis ) I/O per second per disk

26. Response Times and Utilization.

27. Response Time Standard Deviation

28. Concluding Remarks • Parity logging achieves better performance than Raid Level 5 arrays • When data must be preread before being overwritten, Parity Logging is comparable to floating parity and data • Performance is superior to mirroring and floating parity and data when the data to be overwritten is cached

29. Questions • What is parity logging • Describe the general technique of Parity logging. • What is the small write problem, and why it is so important • What are the advantages and disadvantages of floating data and parity