Consistency-preserving Caching of Dynamic Database Content

Transcription

1 Consistency-preserving Caching of Dynamic Database Content Niraj Tolia M. Satyanarayanan Carnegie Mellon University

2 Motivation Database Server Web and App. Servers Easy to geographically distribute web and app. servers Harder to distribute databases Pick Two: Consistency, Availability, Tolerance to Partitions How can you optimize the use of the WAN? 2

3 Outline Motivation Ganesh Overview Design and Implementation Evaluation Conclusion 3

4 High-level Overview of Ganesh Optimizes WAN usage for database accesses Without interpreting queries or results In a database-independent manner Uses a Content Addressable cache to Detect similarity between query results Eliminate redundancy over WAN link Tradeoff increased computation for network savings 4

5 Finding Result Similarity (A 10,000 foot view) SELECT queries 5

6 Finding Result Similarity (A 10,000 foot view) Query 1 SELECT queries 6

7 Finding Result Similarity (A 10,000 foot view) Query 1 SELECT queries 7

8 Finding Result Similarity (A 10,000 foot view) SELECT queries 8

9 Finding Result Similarity (A 10,000 foot view) SELECT queries 9

10 Finding Result Similarity (A 10,000 foot view) SELECT queries 10

11 Finding Result Similarity (A 10,000 foot view) SELECT queries 11

12 Finding Result Similarity (A 10,000 foot view) Query 2 SELECT queries 12

13 Finding Result Similarity (A 10,000 foot view) Query 2 SELECT queries 13

14 Finding Result Similarity (A 10,000 foot view) Data ID SELECT queries 14

15 Finding Result Similarity (A 10,000 foot view) Data ID SELECT queries 15

16 Finding Result Similarity (A 10,000 foot view) SELECT queries 16

17 Finding Result Similarity (A 10,000 foot view) SELECT queries 17

18 Finding Result Similarity (A 10,000 foot view) SELECT queries UPDATE queries are not optimized 18

19 Design Goals Transparent To both the application server and the database Does not weaken consistency Efficiently detects similarity 19

20 Design - Transparency Doesn t require modifications to the application and database server 20

21 Design - Transparency Doesn t require modifications to the application and database server Web and Application Server JDBC Driver DB 21

22 Design - Transparency Doesn t require modifications to the application and database server Web and Application Server JDBC Driver Ganesh Proxy DB 22

23 Design - Transparency Doesn t require modifications to the application and database server Web and Application Server Ganesh Proxy JDBC Driver DB 23

24 Design - Transparency Doesn t require modifications to the application and database server Web and Application Server Ganesh JDBC JDBC Driver Driver Ganesh Proxy JDBC Driver DB 24

25 Proxy-based Architecture Ganesh JDBC Driver Thin but smart -- conservative network use Contains in-memory cache Caches previous results at different granularities Uses a LRU replacement policy Ganesh Proxy Forwards queries to the database Eliminates redundancy in results 25

26 Detecting Similarity Ganesh uses Content Addressability Result Result Hash Cryptographic Hash Hash value is a globally unique identifier Independent of any particular system Infeasible to find another object with same hash If hash values are equal, so are the source objects Any small change in source completely changes hash 26

27 Exploiting Structure Exploit structure in results they look like tables SELECT name, address, zip, FROM USERS Name Address Zip John Doe 412 Avenue Mary Major 821 Lane John Stiles 701 Street T1L1J4 27

28 Exploiting Structure Exploit structure in results they look like tables SELECT name, address, zip, FROM USERS Name Address Zip John Doe 412 Avenue Mary Major 821 Lane John Stiles 701 Street T1L1J4 Result Hash 28

29 Exploiting Structure Exploit structure in results they look like tables SELECT name, address, zip, FROM USERS Name Address Zip John Doe 412 Avenue Mary Major 821 Lane John Stiles 701 Street T1L1J4 Result Hash 29

30 Exploiting Structure Exploit structure in results they look like tables SELECT name, address, zip, FROM USERS Name Address Zip John Doe 412 Avenue Row Hash 1 Mary Major 821 Lane mm@eg.com John Stiles 701 Street T1L1J4 js@eg.com Result Hash 30

31 Exploiting Structure Exploit structure in results they look like tables SELECT name, address, zip, FROM USERS Name Address Zip John Doe 412 Avenue Row Hash 1 Mary Major John Stiles 821 Lane 701 Street T1L1J4 mm@eg.com js@eg.com Row Hash 2 Result Hash 31

32 Exploiting Structure Exploit structure in results they look like tables SELECT name, address, zip, FROM USERS Name Address Zip John Doe 412 Avenue Row Hash 1 Mary Major John Stiles 821 Lane 701 Street T1L1J4 mm@eg.com js@eg.com Row Hash 2 Row Hash 3 Result Hash 32

33 Exploiting Structure Exploit structure in results they look like tables SELECT name, address, zip, FROM USERS Name Address Zip John Doe 412 Avenue Row Hash 1 Mary Major John Stiles 821 Lane 701 Street T1L1J4 mm@eg.com js@eg.com Row Hash 2 Row Hash 3 This process does not interpret data Result Hash 33

34 Putting it all together 34

35 Putting it all together Query 1 35

36 Putting it all together Query 1 36

37 Putting it all together 0x8c3aea3b Query 1 37

38 Putting it all together 0x8c3aea3b Query 1 0x8c3aea3b 38

39 Putting it all together 0x8c3aea3b 39

40 Putting it all together 0x8c3aea3b 40

41 Putting it all together 0x8c3aea3b 41

42 Putting it all together 0xda39a3ee 0x8c3aea3b 42

43 Putting it all together 0xda39a3ee 0x8c3aea3b 43

44 Putting it all together Query 2 0xda39a3ee 0x8c3aea3b 44

45 Putting it all together Query 2 0xda39a3ee 0x8c3aea3b 45

46 Putting it all together 0x8c3aea3b Query 2 0xda39a3ee 0x8c3aea3b 46

47 Putting it all together 0x8c3aea3b Query 2 0xda39a3ee 0x8c3aea3b 47

48 Putting it all together 0x8c3aea3b Query 2 0xda39a3ee 0x8c3aea3b 48

49 Putting it all together 0x8c3aea3b 0xda39a3ee 0x8c3aea3b 49

50 Putting it all together 0x8c3aea3b 0xda39a3ee 0x8c3aea3b 50

51 Putting it all together 0x8c3aea3b 0xda39a3ee 0x8c3aea3b 51

52 Putting it all together 0x8c3aea3b 0xda39a3ee 0x8c3aea3b 52

53 Putting it all together 0x8c3aea3b 0xda39a3ee 0x8c3aea3b 0x8c3aea3b 53

54 Putting it all together 0x8c3aea3b 0xda39a3ee 0x8c3aea3b 0x8c3aea3b No explicit cache-coherence algorithm needed 54

55 Outline Motivation Ganesh Overview Design and Implementation Evaluation Benchmarks and Experimental Setup Results Did exploiting structure help? Conclusion 55

56 Evaluation Benchmarks AUCTION (RUBiS) Models ebay Browsing, bidding, add auctions and feedback, BBOARD (RUBBoS) Models Slashdot Reading articles, adding comments, moderating, Benchmarks have RW and RO variants Performance metrics Throughput (Requests/sec) Latency (Average request response time) 56

57 Experimental Setup Clients Two configurations: Native: Unmodified Setup 57

58 Experimental Setup Clients Ganesh Proxy Two configurations: Native: Unmodified Setup Ganesh: Content-based optimizations used 58

59 Experimental Setup Clients Ganesh Proxy Two configurations: Native: Unmodified Setup Ganesh: Content-based optimizations used Evaluation Parameters 5 Mbit/s + 66ms, 20 Mbit/s + 33ms, 100 Mbit/s 400, 800, 1200, and 1600 clients 59

60 AUCTION Throughput Requests / sec Native Ganesh Higher is better Read-Only Mbit/s 20 Mbit/s 100 Mbit/s Test Clients 60

61 AUCTION - Latency Read-Only 10 1 Native Ganesh Lower is better Mbit/s 20 Mbit/s 100 Mbit/s Test Clients Avg. Resp. Time (sec)

62 BBOARD - Throughput Read-Write Requests / sec Native Ganesh Higher is better Mbit/s 20 Mbit/s 100 Mbit/s Test Clients 62

63 BBOARD - Latency Read-Write Native Ganesh Lower is better Mbit/s 20 Mbit/s 100 Mbit/s Test Clients Avg. Resp. Time (sec)

64 Rabin Fingerprinting vs. Structure Could we have used Rabin fingerprinting? Extensive use in storage systems Chunks data using a stochastic process ` Works well for in-place updates, deletes, insertions Does not work well for query results Reordering of data (ORDER BY) Also hard to pick average chunk size 64

65 Rabin vs. Exploiting Structure BBOARD - Read-Only Mb/s 20 Mb/s 100 Mb/s Test Clients Norm. Throughput

66 Related Work Caching Dynamic Database Content DBCache, DBProxy, MTCache, Per-application consistency model [Gao03, GlobeDB] Backend-scalability [C-JDBC, SSS, Ganeymed] Content-Addressable Systems TCP-level duplicate elimination [Riverbed, Spring00] P2P backup [Pastiche], Storage [Venti, SiS] Distributed File Systems [LBFS, CASPER, PAST] 66

67 Conclusion Ganesh: Optimizes database access over the WAN Transparent Does not weaken consistency Prototype built around Java and the JDBC interface 67