WAN Optimized Replication of Backup Datasets Using Stream-Informed Delta Compression

Transcription

1 WAN Optimized Replication of Backup Datasets Using Stream-Informed Delta Compression Philip Shilane, Mark Huang, Grant Wallace, & Windsor Hsu Backup Recovery Systems Division EMC Corporation

2 Introduction 80% of our customers replicate most of their data off-site for disaster recovery WAN bandwidth often limits throughput Data reduction techniques increase effective throughput Deduplication and local compression are effective Delta compression with stream-informed caching adds 2X additional compression Remote Office Backup System Central Office Backup System

3 Example Of Deduplication And Delta Compression Chunk fp Sketches based on Broder [97 & 00]

4 Example Of Deduplication And Delta Compression Chunk fp sk Maximal Value 1 Maximal Value 2 Maximal Value 3 Maximal Value 4 Sketches based on Broder [97 & 00] super_feature = Rabin_fp(feature 1 feature 4 ) sketch is one or more super_features

5 Example Of Deduplication And Delta Compression Chunk fp sk Chunk fp (duplicate of earlier chunk) Sketches based on Broder [97 & 00] super_feature = Rabin_fp(feature 1 feature 4 ) sketch is one or more super_features

6 Example Of Deduplication And Delta Compression Chunk fp sk Chunk fp sk (similar to earlier chunk) Regions of difference Maximal Value 1 Maximal Value 2 Maximal Value 3 Maximal Value 4 Sketches based on Broder [97 & 00] super_feature = Rabin_fp(feature 1 feature 4 ) sketch is one or more super_features

7 Example Of Deduplication And Delta Compression Chunk fp sk Chunk fp sk (similar to earlier chunk) Regions of difference Transmit fp and differences Sketches based on Broder [97 & 00]

8 Sketch Index Options Full index (simple idea) Requires IO Difficult to update Finds all similarity matches 256 TB capacity 8 KB chunks 16 byte record 0.5 TB index per super-feature

9 Sketch Index Options Full index (simple idea) Requires IO Difficult to update Finds all similarity matches Partial index (slightly better?) Load and evict with LRU policy Not persistent Must be as large as full backup 256 TB capacity 8 KB chunks 16 byte record 0.5 TB index per super-feature

10 Sketch Index Options Partial index would have to hold a full backup 256 to TB be capacity effective 8 KB chunks 16 byte record Full index (simple idea) Requires IO Difficult to update Finds all similarity matches Partial index (slightly better?) Load and evict with LRU policy Not persistent Must be as large as full backup 0.5 TB index per super-feature Stream-informed cache (our contribution) Experimentally demonstrate that delta locality closely matches deduplication locality for backup datasets Updates handled by fingerprint system Little extra memory Finds most similarity matches

11 Sketch Index Options Full index (simple idea) Requires IO Difficult to update Finds all similarity matches Partial index (slightly better?) Load and evict with LRU policy Not persistent Must be as large as full backup 256 TB capacity 8 KB chunks 16 byte record 0.5 TB index per super-feature Partial index has to be large enough to index entire primary storage system

12 Sketch Index Options Full index (simple idea) Requires IO Difficult to update Finds all similarity matches Partial index (slightly better?) Load and evict with LRU policy Not persistent Must be as large as full backup 256 TB capacity 8 KB chunks 16 byte record 0.5 TB index per super-feature Partial index has to be large enough to index entire primary storage system Stream-informed cache (our contribution) Experimentally demonstrate that delta locality closely matches deduplication locality for backup datasets Updates handled by fingerprint system Little extra memory Finds most similarity matches

13 Stream-Informed Locality Similarity search can leverage deduplication locality Sketch cache loaded based on fingerprint cache Fingerprint and Sketch Cache

14 Stream-Informed Locality Similarity search can leverage deduplication locality Sketch cache loaded based on fingerprint cache Full 1 A B C Fingerprint and Sketch Cache fp A fp B fp C sk A sk B sk C

15 Stream-Informed Locality Similarity search can leverage deduplication locality Sketch cache loaded based on fingerprint cache Full 1 A B C Fingerprint and Sketch Cache Container 1 fp A fp B fp C sk A sk B sk C

16 Stream-Informed Locality Similarity search can leverage deduplication locality Sketch cache loaded based on fingerprint cache Full 1 A B C D E F Fingerprint and Sketch Cache Container 1 fp A sk A fp D sk D fp B sk B fp E sk E fp C sk C fp F sk F

17 Stream-Informed Locality Similarity search can leverage deduplication locality Sketch cache loaded based on fingerprint cache Full 1 A B C D E F Fingerprint and Sketch Cache Container 1 Container 2 fp A sk A fp D sk D fp B sk B fp E sk E fp C sk C fp F sk F

18 Stream-Informed Locality Similarity search can leverage deduplication locality Sketch cache loaded based on fingerprint cache Full 1 A B C D E F Store containers to disk Fingerprint and Sketch Cache Container 1 Container 2 fp A sk A fp D sk D fp B sk B fp E sk E fp C sk C fp F sk F

19 Stream-Informed Locality Similarity search can leverage deduplication locality Sketch cache loaded based on fingerprint cache Full 1 Full 2 A A B C D E F Fingerprint and Sketch Cache

20 Stream-Informed Locality Similarity search can leverage deduplication locality Sketch cache loaded based on fingerprint cache Full 1 Full 2 A A B C D E F Load container from disk Fingerprint and Sketch Cache

21 Stream-Informed Locality Similarity search can leverage deduplication locality Sketch cache loaded based on fingerprint cache Full 1 A B C D E F Full 2 A Load container from disk Fingerprint and Sketch Cache Container 1 fp A fp B fp C sk A sk B sk C

22 Stream-Informed Locality Similarity search can leverage deduplication locality Sketch cache loaded based on fingerprint cache Full 1 A B C D E F Full 2 A B C Fingerprint and Sketch Cache Container 1 fp A fp B fp C sk A sk B sk C

23 Stream-Informed Locality Similarity search can leverage deduplication locality Sketch cache loaded based on fingerprint cache Full 1 A B C D E F Full 2 A B C D Load container from disk Fingerprint and Sketch Cache Container 1 fp A fp B fp C sk A sk B sk C

24 Stream-Informed Locality Similarity search can leverage deduplication locality Sketch cache loaded based on fingerprint cache Full 1 A B C D E F Full 2 A B C D E F Fingerprint and Sketch Cache Container 1 Container 2 E is similar to previous chunk E and is a sketch match fp A sk A fp D sk D fp B sk B fp E sk E fp C sk C fp F sk F

25 Replication With Deduplication Remote Office Central Office

26 Replication With Deduplication Send chunk fingerprints Remote Office fp() fp() fp() Central Office

27 Replication With Deduplication Send chunk fingerprints Remote Office fp() fp() fp() Reply with list of missing fingerprints fp() fp() Central Office

28 Replication With Deduplication Send chunk fingerprints Remote Office fp() fp() fp() Reply with list of missing fingerprints fp() fp() Central Office Send missing chunks

29 Replication With Deduplication Send chunk fingerprints Remote Office fp() fp() fp() Reply with list of missing fingerprints fp() fp() Central Office Send missing chunks

30 Replication With Deduplication And Delta Remote Office Central Office m

31 Replication With Deduplication And Delta Remote Office Send chunk fingerprints fp() fp() fp() fp() fp(m) Central Office m

32 Replication With Deduplication And Delta Remote Office Reply with list of missing fingerprints fp() fp(m) Central Office m

33 Replication With Deduplication And Delta Remote Office Central Office Send sketches m sk() sk(m)

34 Replication With Deduplication And Delta Ø m Remote Office Central Office Send sketches m sk() sk(m)

35 Replication With Deduplication And Delta Ø m Remote Office Central Office m Reply with list of fingerprints for similar chunks Ø m

36 Replication With Deduplication And Delta Remote Office Central Office m Send missing chunks and delta encodings () +m

37 Replication With Deduplication And Delta Remote Office Central Office m Send missing chunks and delta encodings m () +m

38 Replication With Deduplication And Delta Remote Office Central Office m Send missing chunks and delta encodings m () +m

39 Replication With Deduplication And Delta Remote Office Central Office m Send missing chunks and delta encodings () +m () +m

40 Replication With Deduplication And Delta Remote Office Central Office m Send missing chunks and delta encodings () +m () () +m +m

41 Replication With Deduplication And Delta Remote Office Central Office m Send missing chunks and delta encodings () +m () () +m +m

42 Replication With Deduplication And Delta Remote Office Central Office m Send missing chunks and delta encodings () +m () +m m

43 Replication With Deduplication And Delta Remote Office Central Office m m Send missing chunks and delta encodings () +m () +m m

44 Properties Of Stream-Informed Delta Compression Pros: Eliminates on-disk sketch index Improves performance fewer disk reads than using a sketch index Small memory footprint Improves compression Cons: Dependent on stream locality and caching to find similar chunks Requires read IO and CPU to process delta chunks

45 Datasets Dataset Type Backup Policy TB Months Source Code Version control repository Weekly full Daily incremental Workstations 16 desktops Weekly full Daily incremental MS Exchange server Daily full System Logs Server s /var directory Weekly full Daily incremental Home Directories Engineers home directories Weekly full Daily incremental

46 Index VS Stream-Informed Cache Cache sized at 12 MB per stream Deduplication 1 super-feature 2 super-features 3 super-features 4 super-features

47 Index VS Stream-Informed Cache Cache sized at 12 MB per stream For one super-feature, compression is within 14% of using an index Deduplication 1 super-feature 2 super-features 3 super-features 4 super-features

48 Index VS Stream-Informed Cache Cache sized at 12 MB per stream For one super-feature, compression is within 14% of using an index Two super-features in a cache is better than an index with one Deduplication 1 super-feature 2 super-features 3 super-features 4 super-features

49 Index VS Stream-Informed Cache Cache sized at 12 MB per stream For one super-feature, compression is within 14% of using an index Two super-features in a cache is better than an index with one Deduplication 1 super-feature 2 super-features 3 super-features 4 super-features

50 Index VS Stream-Informed Cache Cache sized at 12 MB per stream For one super-feature, compression is within 14% of using an index Two super-features in a cache is better than an index with one Deduplication 1 super-feature 2 super-features 3 super-features 4 super-features

51 Index VS Stream-Informed Cache Cache sized at 12 MB per stream For one super-feature, compression is within 14% of using an index Two super-features in a cache is better than an index with one Deduplication 1 super-feature 2 super-features 3 super-features 4 super-features

52 Index VS Stream-Informed Cache Cache sized at 12 MB per stream For one super-feature, compression is within 14% of using an index Two super-features in a cache is better than an index with one 1. Deduplication is slightly different because of different implementations 2. Delta compression in a cache was higher than a full index because of better delta encoding with a candidate in the cache Deduplication 1 super-feature 2 super-features 3 super-features 4 super-features

53 Delta Compression Typical delta improvement is 2X beyond deduplication and GZ compression Choose GZ or Delta with GZ Dataset Deduplication GZ Delta w/ GZ Delta Improv. Source Code 24.9X 7.2X 14.9X 2.1X Workstations 5.7X 2.8X 8.8X 3.1X 6.9X 3.1X 5.8X 1.9X System Logs 57.9X 4.6X 10.2X 2.2X Home Directories 31.7X 3.1X 5.5X 1.8X Compression factors are presented after first week of seeding.

54 Network Throughput Effective throughput is 1-2 orders of magnitude faster

55 Overheads And Limitations Sketches take up about 20 bytes per non-duplicate chunk Uses read IO and CPU on source and destination Sketching is a 20% slowdown on writes, but only for non-duplicates Scales linearly at destination with number of streams

56 Overheads And Limitations Sketches take up about 20 bytes per non-duplicate chunk Resource utilization during replication Uses read IO and CPU on source and destination Sketching is a 20% slowdown on writes, but only for non-duplicates Scales linearly at destination with number of streams Shared sketch cache affects compression System sized to handle 20 streams Number of systems replicating to a single destination With 25 streams compression loss of 0-12% With 50 streams compression loss of 0-27%

57 Overheads And Limitations Sketches take up about 20 bytes per non-duplicate chunk Uses read IO and CPU on source and destination Sketching is a 20% slowdown on writes, but only for non-duplicates Scales linearly at destination with number of streams Shared sketch cache affects compression System sized to handle 20 streams With 25 streams compression loss of 0-12% With 50 streams compression loss of 0-27%

58 Customer Results Median customer has 2X delta compression beyond deduplication

59 Related Work Optimized network transfer Spring00, Muthitacharoen01, Eshghi07, and Park07 File synchronization Tridgell00, Suel04 Delta compression Burns97, Mogul97, Hunt98, Chan99, MacDonald00, Suel02, Trendafilov02, and Chen04 Similarity detection Brin94, Manber94, Broder[97 & 00], Douglis03, Kulkarni04, You04, Jain05, and Aronovich09 Deduplicated storage Policroniades04, and Bobbarjung06 Stream-informed deduplication Zhu08, Bhagwat09, Lillibridge09, Min10, Guo11, and Xia11

60 Conclusion Delta locality closely matches deduplication locality for backup datasets Good scalability Stream-informed delta compression is effective with a small cache CPU and IO utilization is low Product allows customers to replicate and protect twice as much data across a WAN

61 Questions?

62