TRANSACTIONAL FLASH CARSTEN WEINHOLD. Vijayan Prabhakaran, Thomas L. Rodeheffer, Lidong Zhou

Transcription

1 Department of Computer Science Institute for System Architecture, Operating Systems Group TRANSACTIONAL FLASH Vijayan Prabhakaran, Thomas L. Rodeheffer, Lidong Zhou CARSTEN WEINHOLD

2 MOTIVATION Transactions have proven useful: File systems Databases systems Common approaches: Copy on write Write ahead logging Hard to get right Everybody reinvents the wheel... 2

3 BETTER DISKS? Transaction support could be built into disk Much simpler file systems / databases Problem: Copy on write causes fragmentation Slow seeks needed when reading Solutions: Reorganize data in cleaning process Checkpointing + update home location 3

4 FLASH BASICS Typical solid state disk: Controller + multiple flash packages Small mount of RAM to buffer I/O requests, internal data structures Data organization: Packages contain planes, blocks, pages 128 bytes of metadata for each 4 KB page Spare memory for data from damaged areas 4

5 HOW IT WORKS Random reads / writes are fast Overwriting is slow: Entire block must be erased Takes in the order of milliseconds Limited number of erase / write cycles Out-of-place updates avoid overwriting Garbage collection reclaims old pages Wear leveling minimizes per-block erasure 5

6 TXFLASH!"#$ %&' (%)%*%+$,-+)$.+!"#$%&$'(#)* &+'"$*,%-.*!"#$% 7$.%9 :%*#$!5$$ *#03;+ :< =$)%'%)% 7$308$5-102"3 /0..") 102"3!:1 6%5*%2$ /0##$3)05 4+0#%)"0& 1%-$5 :<!#%+> /0&)50##$5!#%+>?%3;%2$+ 6

7 TXFLASH TxFlash builds on top of existing flash storage controllers Introduces additional command: WriteAtomic(p0, p1,... pn-1) 7

8 REDO LOGGING Most file systems use redo logging: Intention records written to storage: Pages with data Metadata describing location, etc. Extra write for commit record, after intention records are persistent Data from log copied to home locations in checkpoint process Recovery: redo committed transactions 8

9 TXN PROTOCOLS /"3 ;"$" 5"<). "-* =),#(&-. :,"-#"%$(&- A; :,"*($(&-"8 6&''($ /03 ;"$" 5"<). "-* =),#(&-. >)?$ 5"<). "-* =),#(&-. 9('28) 67%8(% 6&''($ /%3 ;"$" 5"<). "-* =),#(&-. >)?$ 5"<). "-* =),#(&-. %&''($$)* +),#(&-. /0"%1 2&(-$),3 4"%1 5&(-$), 67%8(% 6&''($ 9

10 CYCLIC COMMIT Requirement: data + metadata can be stored together efficiently No extra write for commit record: Each intention record has next link Last intention record points to first one Concurrent writes possible for all records Recovery: full cycle in storage describes committed transaction 10

11 SCC!" #" $%&' (" 4" )'*+,-../01'* 2" 3" */22.::); +),-./0 <08/22.::); +),-./0 */22.::); '0; (',3'() 8/==)8:); +),-./0 A.--.0( +),-./0 #'89 &/.0:), 11

12 SCC VS. BPCC Simple Cyclic Commit: No overlapping transactions (isolation) Uncommited intention records must be erased before starting new transaction Back Pointer Cyclic Commit: Extra back pointer: last committed transaction Avoids erase cycle after aborted transaction More complex garbage collection / recovery 12

13 BPCC &'()!" #" *" $" %" 4" +),-./0 0123), 5" 6" 7" */22.::); +),-./0 <08/22.::); +),-./0 */22.::); '0; (',3'() 8/==)8:); +),-./0 A.--.0( +),-./0 #'89 &/.0:), 13

14 EVALUATION 14

15 TXN SIZE Figure 6: Impact of Transaction Size. Transaction throughput vs. 15

16 ABORTED TXNS 16

17 PERFORMANCE 17

18 COSTS TxFlash SSD +TC +SCC +BPCC LOC : Implementation Complexity. Lines of code in S 18

19 DISCUSSION Model-checked! Cool! Is the cyclic commit protocol really new? File systems do not cancel transactions. Do we really need BPCC? Databases cancel transactions, but TxFlash is not fit for them yet. Will BPCC still suffice? Are page writes atomic? 19

20 BACKUP SLIDES 20

21 ABORTED TXNS $%&'!" #" (" $%&'!" #" (" )'*+,-../01'* 2" 3" (%+' 4%5 (%+' 415 (-00,88'9 )'*+,-. :.7-00,88'9 )'*+,-. <'=8>;,.? (-00,88'9 %.9 &%*1%&' 7-;;'78'9 )'*+,-. 6*%.+%78,-. %1-*8 21