Sang-Phil Lim Sungkyunkwan University. Sang-Won Lee Sungkyunkwan University. Bongki Moon University of Arizona

Transcription

1 Sang-Phil Lim Sungkyunkwan University Sang-Won Lee Sungkyunkwan University Bongki Moon University of Arizona

2 Table of Contents Mo.va.on Background NAND Flash Memory and Flash Transla>on Layer (FTL) FAST FTL and OLTP Write Pa?erns FASTer FTL for OLTP Applica.ons Performance Evalua.on Conclusion and Future Work

3 Mo.va.on FAST FTL [Sang- Won Lee et al, ACM TECS 07] Originally designed for random writes FAST has been cri.cized [DFTL: Aayush Gupta et al, ASPLOS 09, LAST: Sungjin Lee et al, SPEED 08] With 3% log space, performance and fluctua>on No special mechanism for hot/cold separa>on A large scale flash SSD For be\er performance, it can employ larger log space Revisit FAST with OLTP workloads Cost compe>>veness

4 Background : NAND Flash Memory NAND Flash memory organiza.on & chip- level performance Flash Chip Block 0 Limita.ons Block 1 Page 0 Page 1 Page m- 1 Block n- 1 Erase- before- write : No in- place update Data can only be wri\en sequen>ally Block wears out a`er 100K erases Media NAND Flash [1] HDD [2] Access latency Read Write Erase 25us (2KB) 12.7ms (2KB) 200us (2KB) 13.7ms (2KB) 1,500us (128KB) [1] : Samsung Large- block SLC NAND (K9WAG08U1A) [2] : Seagate 7200rpm Hard Disk (ST380011A) -

5 Background : NAND Flash Memory Erase- before- write limita.on Flash Memory New write (2KB) : 0.2ms Overwrite (2KB) 63 page copy- backs 1 new page write 1 old block erase 16ms ( > 80x new page write)

6 Background : Flash Transla.on Layer(FTL) A sotware layer that allows the flash memory to look like a HDD Address mapping : logical to physical Garbage collec>on & power- off recovery Wear- leveling & bad block management etc. Read Read File System Mismatch Write Write Erase Low- level Device Driver Read File System FTL Write Low- level Device Driver Popular FTL algorithms FMAX, BAST, FAST, Super block, LAST DFTL, DAC, etc... Flash memory Flash memory

7 FAST FTL A popular log- based FTL [Sang- Won Lee et al, ACM TECS 07] FAST FTL is designed for small random writes Sectors from one logical block can be mapped to any log block Flash Memory Original Data Blocks Log Blocks

8 FAST FTL vs. Temporal Locality FAST FTL can handle temporal locality without any overhead f or iden>fying hot/cold- ness Free block becomes a new log block Invalidate old data Merge for the vali d pages vic>m Log area : Hot page : Warm page : Cold page : Invalid page : valid page Merge opera.on (individual) Search all valid pages 64 page copy- backs (to free block) Replace the free block to new data block 1 old data block erase 16ms ( > 80x new page write)

9 Write Pa?erns in OLTP applica.ons Randomly sca\ered over a large address space Write skewed : non- uniformly distributed access frequencies

10 Impact of Write Intervals on FTLs We classified the pages using the concept of write interval Hot/Warm/Cold page Temporal locality in OLTP workload OLTP write pa\erns may match well for FAST FTL Write interval vs. log window size? 500K 200K 100K

11 Cri.cism of FAST FTL FAST FTL is cri.cized in DFTL [Aayush Gupta et al, ASPLOS `09] They said... FAST dose not provide any special mechanism to handle tempor al locality in random streams. With 3% log space, FAST shows poor performance and high varia>on

12 Impact of log window size in FAST FTL With larger log space, FAST can exploit temporal locality! Trade- off : manufacturing cost vs. throughput Low merge cost High merge cost

13 FASTer FTL : Introduc.on FASTer FTL for OLTP workloads A new FTL scheme which is enhancement of FAST FTL Be\er performance than FAST Uniform response >me Main key ideas Second chance policy Isola>on area

14 FASTer FTL : Key ideas 1) Second chance policy Free block becomes a new log block vic>m Log area : Hot page : Warm page : Cold page : Invalid page Con>nue to new writes Give a second chance, i nstead of merging Cons. 3 page copy- backs (decrease log block u>liza>on) We can get more chances to skip merge op. for the warm pages Free block Pros. (= skip merges for the warm pages) Avoid 2 x (64 page copy- backs) Avoid 2 old data block erases : valid page Pros. >> Cons.

15 FASTer FTL : Key ideas 1) Second chance policy Give the second chance for the warm pages to be invalidated Exploit temporal locality more by doubled log window Pros. & Cons. of second chance policy But, Pros. >> Cons.

16 FASTer FTL : Key ideas 2) Isola.on area Free block becomes a new log block Write buffering for cold pages Mi>gate response >me fluctua>on vic>m Log area Isola.on area : Already received second chance : Not received second chance yet : invalidate page Con>nue to new writes Free block Merge one cold dirty page when a new page write to the log block

17 Performance Evalua.on FASTer shows be?er throughput than others Outperformed FAST by more than 30 percent in elapsed >me Even similar with page- level mapping FTL [A. Kawaguchi et al, TCON 95]

18 Performance Evalua.on FASTer also shows uniform response.me

19 Performance Evalua.on Effect of write skewedness degree c.f) (x/y) means x% of writes are directed to y% of pages.

20 Conclusion and Future Work Recent trends in NAND technology have made SSDs more viab le in the enterprise storage market In this paper, we proposed FASTer FTL as an enhancement of t he FAST FTL In the future, we will explain FASTer FTL more theore>cally an d evaluate with various real workloads

21 Thank you for your a?en.on Ques.ons & Answers