Resolving Journaling of Journal Anomaly via Weaving Recovery - PowerPoint PPT Presentation

NVRAMOS ‘14 10.30. 2014 Resolving Journaling of Journal Anomaly via Weaving Recovery Information into DB Page Beomseok Nam UNIST

Outline Motivation  • Journaling of Journal Anomaly How to resolve Journaling of Journal anomaly  • Multi-Version B-Tree (MVBT) Optimizations of Multi-Version B-Tree  • Lazy Split • Reserved Buffer Space • Lazy Garbage Collection • Metadata Embedding • Disabling Sibling Redistribution Evaluation  Conclusion  2

Storage I/O Problems in Android  Performance Bottleneck  Lifetime of Storage 1 1 0 0 1 0 0 1 0 1 1 0 0 1 0 1 0 1 0 1 0 Cause = Excessive IO 3

Android I/O Stack Apps SQLite Txn Journaling Insert/Update/Delete/Select Misaligned Interaction EXT4 Write() Journaling Read/Write Block Device Driver 4

Journaling of Journal Anomaly  Journaling in SQLite (TRUNCATE mode) Insert a DB entry SQLite Open rollback journal. Record the data to journal. Put commit mark to journal . Insert entry to DB Truncate journal . 5

EXT4 Journaling (ordered mode) SQLite write() EXT4 Write data block Write EXT4 journal Write journal metadata Write journal commit 6

Journaling of Journal Anomaly insert SQLite One insert of 100 Byte  9 Random Writes of 4KByte EXT4 Write data block Write EXT4 journal Write journal metadata Write journal commit 7

How to Resolve Journaling of Journal? Database Insertion Database Journaling Insert a DB entry SQLite EXT4 Journaling of journal anomaly. 8

How to Resolve Journaling of Journal? fsync() fsync() fsync() Journal DB Journal+DB = Version-based B-Tree (MVBT) 9

Version-based B-Tree (MVBT)  Versioning Insert 10 Update 10 with 20 Time T1 T2 10 [T1, ∞ ) 10 [T1, T2) Dead Entry 20 [T2, ∞ ) Do not overwrite old version data Do not need a rollback journal 10

Node Split in Multi-Version B-Tree P1 5 [3~ ∞ ) key 25, ver=5 10 [2~ ∞ ) 12 [4~ ∞ ) 40 [1~ ∞ ) 11

Node Split in Multi-Version B-Tree Dead P1 Node 5 [3~5) key 25, ver=5 10 [2~5) 12 [4~5) 40 [1~5) 12

Node Split in Multi-Version B-Tree P1 5 [3~5) key 25, ver=5 10 [2~5) 12 [4~5) 40 [1~5) 13

Node Split in Multi-Version B-Tree key 25, ver=5 P3 P2 P1 5 [5~ ∞ ) 5 [3~5) 10 [5~ ∞ ) 10 [2~5) 12 [5~ ∞ ) 12 [4~5) 40 [5~ ∞ ) 40 [1~5) 14

Node Split in Multi-Version B-Tree P4 10 [5~ ∞ ) : P3 ∞ [5~ ∞ ) : P2 ∞ [0~5) : P1 key 25, ver=5 P3 P1 P2 5 [5~ ∞ ) 5 [3~5) 10 [5~ ∞ ) 10 [2~5) 12 [5~ ∞ ) 12 [4~5) 25 [5~ ∞ ) 40 [1~5) 40 [5~ ∞ ) 15

Node Split in Multi-Version B-Tree One more dirty page than P4 dirty original B-Tree 10 [5~ ∞ ) : P3 ∞ [5~ ∞ ) : P2 ∞ [0~5) : P1 P3 P1 P2 dirty dirty dirty 5 [5~ ∞ ) 5 [3~5) 10 [5~ ∞ ) 10 [2~5) 12 [5~ ∞ ) 12 [4~5) 25 [5~ ∞ ) 40 [1~5) 40 [5~ ∞ ) 16

I/O Traffic in MVBT DB buffer cache fsync() MVBT Reduce the number of dirty pages! 17

I/O Traffic in MVBT DB buffer cache fsync() MVBT DB buffer cache fsync() LS -MVBT 18

Optimizations in Android I/O >> t read t write Read Write Multiple Write Insert Transaction 1 Write Transaction 2 Single Insert Transaction Write Transaction 3 19

Lazy Split Multi-Version B-Tree (LS-MVBT)  Optimizations Lazy Split Disabling Reserved Sibling Buffer Space Redistribution LS-MVBT Lazy Garbage Metadata Collection Embedding 20

Optimization1: Lazy Split  Legacy Split in MVBT P4 dirty 10 [5~ ∞ ) : P3 ∞ [5~ ∞ ) : P2 ∞ [0~5) : P1 4 dirty pages P3 P1 P2 dirty dirty dirty 5 [5~ ∞ ) 5 [3~5) 10 [5~ ∞ ) 10 [2~5) 25 [5~ ∞ ) 12 [4~5) 12 [5~ ∞ ) 40 [1~5) 40 [5~ ∞ ) 21

Optimization1: Lazy Split P1 5 [3~ ∞ ) key 25, ver=5 10 [2~ ∞ ) 12 [4~ ∞ ) 40 [1~ ∞ ) 22

Optimization1: Lazy Split Lazy Node: Half-dead, P1 Half-live 5 [3~ ∞ ) key 25, ver=5 10 [2~ ∞ ) 12 [4~5 ) 40 [1~5 ) 23

Optimization1: Lazy Split P1 5 [3~ ∞ ) key 25, ver=5 10 [2~ ∞ ) 12 [4~5 ) 40 [1~5 ) 24

Optimization1: Lazy Split key 25, ver=5 P2 P1 5 [3~ ∞ ) 10 [2~ ∞ ) 12 [5~ ∞ ) 12 [4~5 ) 40 [5~ ∞ ) 40 [1~5 ) 25

Optimization1: Lazy Split P3 10 [5~ ∞ ) : P1 ∞ [0~5 ) : P1 ∞ [5 ~ ∞ ) : P2 key 25, ver=5 P1 P2 5 [3~ ∞ ) 12 [5~ ∞ ) 10 [2~ ∞ ) 12 [4~5 ) 25 [5~ ∞ ) 40 [1~5 ) 40 [5~ ∞ ) 26

Optimization1: Lazy Split  Lazy Node Overflow P3 10 [5~ ∞ ) : P1 ∞ [0~5 ) : P1 ∞ [5 ~ ∞ ) : P2 key 8, ver = 6 P1 P2 5 [3~ ∞ ) 10 [2~ ∞ ) 12 [5~ ∞ ) Dead Dead 25 [5~ ∞ ) 12 [4~5 ) entries entries 40 [1~5 ) 40 [5~ ∞ ) 27

Optimization1: Lazy Split  Lazy Node Overflow → Garbage collect dead entries P3 10 [5~ ∞ ) : P1 ∞ [0~5 ) : P1 ∞ [5 ~ ∞ ) : P2 key 8, ver = 6 P1 P2 5 [3~ ∞ ) 10 [2~ ∞ ) 12 [5~ ∞ ) 25 [5~ ∞ ) 12 [4~5 ) 40 [1~5 ) 40 [5~ ∞ ) 28

Optimization1: Lazy Split  Lazy Node Overflow → Garbage collect dead entries P3 10 [5~ ∞ ) : P1 ∞ [0~5 ) : P1 ∞ [5 ~ ∞ ) : P2 key 8, ver = 6 P1 P2 5 [3~ ∞ ) 8 [6~ ∞ ) 12 [5~ ∞ ) 25 [5~ ∞ ) 10 [2~ ∞ ) 40 [5~ ∞ ) But, what if the dead entries are being accessed by other transactions? 29

Optimization2: Reserved Buffer Space  Option 1. Wait for read transactions to finish P3 10 [5~ ∞ ) : P1 ∞ [0~5 ) : P1 ∞ [5 ~ ∞ ) : P2 key 8, ver = 6 P1 P2 5 [3~ ∞ ) 12 [5~ ∞ ) 10 [2~ ∞ ) 25 [5~ ∞ ) 12 [4~5 ) 40 [1~5 ) 40 [5~ ∞ ) 30

Optimization2: Reserved Buffer Space  Option 2. Split as in legacy MVBT split P3 10 [5~ ∞ ) : P1 ∞ [0~5 ) : P1 ∞ [5 ~ ∞ ) : P2 key 8, ver = 6 P1 P2 P3 5 [3~6) 5 [6~ ∞ ) 12 [5~ ∞ ) 1 0 [2~6) 10 [6~ ∞ ) 12 [4~5 ) 25 [5~ ∞ ) 40 [1~5 ) 40 [5~ ∞ ) 31

Optimization2: Reserved Buffer Space  Option 2. Split as in legacy MVBT split P3 10 [5~6 ) : P1 ∞ [0~5 ) : P1 ∞ [5 ~ ∞ ) : P2 10 [6 ~ ∞ ) : P3 key 8, ver = 6 P3 P1 P2 5 [6~ ∞ ) 5 [3 ~6) 12 [5~ ∞ ) 8 [6~ ∞ ) 10 [2 ~6) 25 [5~ ∞ ) 10 [6~ ∞ ) 12 [4~5 ) 40 [1~5 ) 40 [5~ ∞ ) 32

Optimization2: Reserved Buffer Space  Option 3. Pad some buffer space in tree nodes P3 10 [5~ ∞ ) : P1 ∞ [0~5 ) : P1 ∞ [5 ~ ∞ ) : P2 key 8, ver = 6 P1 P2 12 [5~ ∞ ) 10 [2~ ∞ ) Buffer space is used 25 [5~ ∞ ) 12 [4~5 ) when lazy node is full 40 [1~5 ) 40 [5~ ∞ ) If buffer space is also full, 9 [6~ ∞ ) split as in legacy MVBT 33

Rollback in LS-MVBT Similar to rollback in MVBT • P3 P3 P3 P3 Rollback ∞ [0~ ∞ ) : P1 ∞ [0~5 ) : P1 10 [5~ ∞ ) : P1 10 [5~ ∞ ) : P1 ∞ [0~5 ) : P1 ∞ [0~5 ) : P1 Txn 5 ∞ [5 ~ ∞ ) : P2 ∞ [5 ~ ∞ ) : P2 crashes Number of dirty nodes touched by rollback of LS- MVBT is also smaller than that of MVBT. P1 P1 P2 P1 P1 5 [3~ ∞ ) 5 [3~ ∞ ) 12 [5~ ∞ ) 5 [3~ ∞ ) 5 [3~ ∞ ) 10 [2~ ∞ ) 10 [2~ ∞ ) 25 [5~ ∞ ) 10 [2~ ∞ ) 10 [2~ ∞ ) 12 [4~5 ) 12 [4~5 ) 40 [5~ ∞ ) 12 [4~5 ) 12 [4~ ∞ ) 40 [1~5 ) 40 [1~5 ) 40 [1~ ∞ ) 40 [1~5 ) 34

Optimization3: Lazy Garbage Collection  Periodic Garbage Collection P1 Transaction 1 P3 P1 Transaction buffer cache P2 35

Optimization3: Lazy Garbage Collection  Periodic Garbage Collection Transaction 1 Stopped P3 P1 Transaction buffer cache P2 P1 P2 P3 fsync() Garbage Extra Dirty Pages Collection GC buffer cache 36

Optimization3: Lazy Garbage Collection  Lazy Garbage Collection Do not garabge collect if no space is needed P1 fsync() Transaction 1 Insert P3 P1 Transaction buffer cache P2 No Extra Dirty Page by GC 37

Optimization4: Metadata embedding  Version = “File Change Counter” in DB header page Header Page 1 dirty 6 5 ++ Write 6 Transaction Page 2 Commit 2 dirty pages ... Read 5 Transaction dirty Page N 15 [ 6 ~ ∞ ) 25 [5~ ∞ ) 40 [1~ ∞ ) 38

Optimization4: Metadata embedding  Flush “File Change Counter” to the last modified page and RAMDISK Header Page 1 Write 6 Transaction Page 2 4 Commit ... Read 5 Transaction Page N 5 6 RAMDISK 15 [ 6 ~ ∞ ) 25 [5~ ∞ ) 5 6 40 [1~ ∞ ) 39

Optimization4: Metadata embedding  Flush “File Change Counter” to the last modified page and RAMDISK Header Page 1 Write 6 Transaction Page 2 4 Commit CRASH ... Read 5 Transaction Page N 6 RAMDISK dirty 15 [ 6 ~ ∞ ) 25 [5~ ∞ ) 6 6 40 [1~ ∞ ) 1 dirty page Volatile 40

Optimization5: Disable Sibling Redistribution  Sibling redistribution hurts insertion performance dirty P5: P5: 10 : P4 10 : P4 10 12 Insertion of key 20 40 : P3 40 : P3 40 25 : P2 : P2 ∞ ∞ 4 dirty pages P4: dirty dirty P2: dirty P3: 12 55 5 15 20 10 ∞ 25 40 • Disable sibling redistribution → avoid dirtying 4 pages 41

Optimization5: Disable Sibling Redistribution  Disabled sibling redistribution P5 : dirty 10 : P4 40 : P3 15 : P5 Insertion of key 20 90 : P2 P3: P5: P2: P4: dirty dirty 20 12 55 5 90 15 10 25 40 3 dirty pages 42

Summary  Optimizations Lazy Split Avoid dirtying an extra page when split occurs Disabling Reserved Sibling Buffer Space Redistribution Reduce the probability LS-MVBT Do not touch siblings to of node split make search faster Lazy Garbage Metadata Collection Embedding Delete dead entries on the Avoid dirtying header page next mutation 43