Increasing Buffer-Locality for Multiple Index Based Scans through Intelligent Placement and Index Scan Speed Control

Transcription

1 IM Research Increasing uffer-locality for Multiple Index ased Scans through Intelligent Placement and Index Scan Speed Control Christian A. Lang ishwaranjan hattacharjee Tim Malkemus Database Research Group IM T.J. Watson Research Center Kwai Wong IM Toronto Lab

2 Goal Improve query performance (throughput+latency) for ad-hoc index scan-heavy multi-query workloads (e.g., DSS workloads) with minimal architecture dependency/impact 2

3 Example DSS Queries select sum(l_extendedprice*l_discount) as revenue from lineitem where l_shipdate >= 01/01/2006 and l_shipdate < 01/01/2006' + interval '1' year and l_quantity > 10; select sum(l_extendedprice*l_discount) as revenue, avg(l_extendedprice*l_discount) as avgsale from lineitem where l_shipdate >= 10/01/2006 and l_shipdate < 10/01/ interval 3' month and l_quantity > 30; 3

4 Challenges in Multi-Query DSS Workloads DSS workloads include scanning of large amounts of data (e.g., aggregate calculation) Cannot optimize ahead of time (many ad-hoc queries, unknown start times) Trend: even more I/O bound queries (disk seek/access times not keeping up with capacity growth/cpu speed) Sub-optimal cache reuse (current RDMS treat queries (mostly) in isolation) 4

5 Known Solutions (not to scale) Can handle drift/ad-hoc queries? RT-DMS P-aware QO [Ramamurthy/DeWitt05] Yes Cooperative Scans [Zukowski07] QPipe [Harizopoulos05] TEM [Kotidis01] SISCANs (cache and index independent) ARC LRU-K LRFU Teradata? SQLServer? Oracle? No (static queries) NonStop SQL/MX [Clear99] Multi-query optimization Higher Lower Impact on existing architecture 5

6 Outline Current Index Scan Architecture SISCAN Circular Index Scan Placement Speed Control Implementation Issues Index-independent Relative SISCAN Location ufferpool-independent SISCAN-aware Caching Experimental Results Conclusions 6

7 Current Index Scan Architecture ufferpool Table pages stored on disk Index structure 7

8 Current Index Scan Architecture HJ NLJ HJ IXSCAN LINEITEM NLJ IXSCAN CUSTOMER Query 1 Tscan ORDERS ufferpool Table pages stored on disk Index structure 8

9 Current Index Scan Architecture HJ NLJ HJ Query Execution IXSCAN LINEITEM NLJ IXSCAN CUSTOMER Index scan process A Query 1 Tscan ORDERS ufferpool Table pages stored on disk Index structure 9

10 Current Index Scan Architecture A 1 ufferpool 1 1 Table pages stored on disk Index structure 10

11 Current Index Scan Architecture A ufferpool Table pages stored on disk Index structure 11

12 Current Index Scan Architecture HJ NLJ HJ Query Execution Tscan ORDERS Query 2 IXSCAN CUSTOMER NLJ IXSCAN CUSTOMER A Index scan process ufferpool Table pages stored on disk Index structure 12

13 Current Index Scan Architecture A ufferpool Table pages stored on disk Index structure 13

14 Current Index Scan Architecture A ufferpool Table pages stored on disk Index structure 14

15 Current Index Scan Architecture A Pages read by A have to be re-read by extra I/O, slowdown ufferpool Table pages stored on disk Index structure 15

16 Outline Current Index Scan Architecture SISCAN Circular Index Scan Placement Speed Control Implementation Issues Index-independent Relative SISCAN Location ufferpool-independent SISCAN-aware Caching Experimental Results Conclusions 16

17 SISCAN Circular Index Scan A ufferpool Table pages stored on disk Index structure 17

18 SISCAN Circular Index Scan Start scan at A s key position bufferpool pages reused A ufferpool Table pages stored on disk Index structure 18

19 SISCAN Circular Index Scan When reaches end key wrap around to start key A ufferpool Table pages stored on disk Index structure 19

20 SISCAN Circular Index Scan After wrapping, finishes the remaining key range break single IXSCAN into two ufferpool Table pages stored on disk Index structure 20

21 Placement: Where to Start with Multiple Active SISCANs? Scan C Current location Scan A Current location Scan Current location Scan ranges Scan A Scan Scan C 21

22 Placement: Where to Start with Multiple Active SISCANs? Scan C Current location Scan A Current location Scan Current location Scan ranges Scan A Scan Scan C New scan Where to start new scan? need more information 22

23 Placement: Where to Start with Multiple Active SISCANs? Scan C Current location Scan A Current location Scan Current location Scan ranges Scan A Scan Scan C New scan Scan A Scan C Scan location Current time time 23

24 Placement: Where to Start with Multiple Active SISCANs? Scan C Current location Scan A Current location Scan Current location Scan ranges Scan A Scan Scan C New scan Scan A Scan C Scan location Current time New scan starts with time 24

25 Estimating Sharing Potential Current time C A Scan key range for E loc D time 25

26 Estimating Sharing Potential Current time C E A Scan key range for E loc D time 26

27 Estimating Sharing Potential Current time C E A Scan key range for E loc D time Number of reads: 3 *15 27

28 Estimating Sharing Potential Current time C E A Scan key range for E loc D time Number of reads: 3 + *15 1 *30 28

29 Estimating Sharing Potential Current time C E A Scan key range for E loc D time Number of reads: *15 *30 2 *15 29

30 Estimating Sharing Potential Current time C E A Scan key range for E loc D time Number of reads: *15 *30 *15 *20 30

31 Estimating Sharing Potential Current time C E A Scan key range for E loc D time Number of reads: = 195 *15 *30 *15 *20 *10 31

32 Estimating Sharing Potential Current time C Scan key range for E A E loc D time 32

33 Estimating Sharing Potential Current time C Scan key range for E A E loc D time Number of reads: 2 *15 33

34 Estimating Sharing Potential Current time C Scan key range for E A E loc D time Number of reads: *15 *20 34

35 Estimating Sharing Potential Current time C Scan key range for E A E loc D time Number of reads: *15 *20 2 *40 35

36 Estimating Sharing Potential Current time C Scan key range for E A E loc D time Number of reads: = 180 *15 *20 *40 *15 36

37 Estimating Sharing Potential Scan key range for E loc A Current time D C E Starting E near A is the better choice time Number of reads: = 180 *15 *20 *40 *15 37

38 Problem solved? No, scans drift apart! location LINEITEM scans during TPC-H 38 time

39 Problem with Drift Key order A C D E F Initial reader of pages in key order E E E F F 39

40 Problem with Drift Key order A C D E F Initial reader of pages in key order E E E F F 3 disk reads 40

41 Problem with Drift Key order A C D E F Initial reader of pages in key order D D E E F F 4 disk reads Drift leads to extra disk reads; How to tolerate some drift without being too rigorous? 41

42 SISCAN Speed Control Key order A C D E F Pages in key order Group 1 Group 2 Group 3 Footprint bufferpool size Greedy algorithm: group nearby SISCANs until sum of group footprint exceeds bufferpool size 42

43 SISCAN Speed Control Key order A C D E F Pages in key order Group 1 Group 2 Group 3 Leader (wait for A if necessary) Leader (wait for C and D if necessary) Leader (no other group members) 43 Delay leader until group size bufferpool size / #groups Upper bound on wait time per SISCAN; Details similar to throttling for table scan sharing [Lang et al., ICDE 07]

44 Outline Current Index Scan Architecture SISCAN Circular Index Scan Placement Speed Control Implementation Issues Index-independent Relative SISCAN Location ufferpool-independent SISCAN-aware Caching Experimental Results Conclusions 44

45 Index-Independent Relative SISCAN Locations Problem: hard to determine relative IXSCAN locations (while leaving the index a black box ) Example (key, page): ( Alice, 12), ( ob, 38), ( ob, 91), ( Carol, 2) What are the relative locations of these scans? How far apart are they? 45

46 Index-Independent Relative SISCAN Locations A s Anchor A 46

47 Index-Independent Relative SISCAN Locations A s Anchor A s offset A 47

48 Index-Independent Relative SISCAN Locations s Anchor A s Anchor A s offset A 48

49 Index-Independent Relative SISCAN Locations s Anchor A s Anchor s offset A s offset A Distance between A and unknown 49

50 Index-Independent Relative SISCAN Locations s Anchor A s Anchor s offset A s offset A 50

51 Index-Independent Relative SISCAN Locations A s Anchor s Anchor s offset A s offset A Distance between A and = A s offset s offset 51

52 Index-Independent Relative SISCAN Locations Anchor Anchor offsets offsets A C D E Partial ordering between SISCANs without details of the index structure 52

53 ufferpool-independent SISCAN-aware Caching Key order A C D E F Pages in key order Group 1 Group 2 Group 3 53

54 ufferpool-independent SISCAN-aware Caching Key order A C D E F Pages in key order Group 1 Group 2 Group 3 Trailer (mark page as LOW priority) Leader (mark page as HIGH priority) LOW priority HIGH priority LOW priority Don t need to change caching algorithm; need only LOW/HIGH priority hints; Details similar to [Lang et al., ICDE 07] 54

55 Architectural Changes scan index I from startkey to endkey Index scan process Index scan manager get next page id Index 55

56 Architectural Changes New component scan index I from startkey to endkey Minor modifications Index scan process Untouched Index scan manager get next page id Index 56

57 Outline Current Index Scan Architecture SISCAN Circular Index Scan Placement Speed Control Implementation Issues Index-independent Relative SISCAN Location ufferpool-independent SISCAN-aware Caching Experimental Results Conclusions 57

58 Experimental Results Setup Platforms: 1. HP Integrity rx5670 (4 Itanium2 proc/1ghz, HP-UX, 15G, FAStT) 2. 8-node p660 cluster (4 PowerPC/600MHz, AIX, 8G, 16 SSA disks) 100G TPC-H database ufferpool size 5% of D size Standard MDC indexes / no hand-tuning 58

59 Staggered Q6 (I/O intensive) CPU Usage Stats For 3 Steams 3 Streams Timings ase SS ase SS % Of Total Time Time 0 User System Idle Wait 1st Q6 2nd Q6 3rd Q6 I/O wait reduced by 50%; More than 50% gain in response time 59

60 Staggered Q1 (CPU intensive) CPU Usage Stats Query Timings ase ScanShare ase Scan Share % Of Total Time Timings in Seconds 20 0 User System Idle Wait 1st Q1 2nd Q1 3rd Q1 Noticeable reduction in response time even for CPU bound queries 60

61 TPC-H Throughput: Per-stream Gains time (s) base ss stream 20% reduction in response time for all streams 61

62 TPC-H Throughput: Per-Query Gains base ss Q1 Q10 Q11 Q12 Q13 Q14 Q15a Q16 Q17 Q18 Q19 Q2 Q20 Q21 Q22 Q3 Q4 Q5 Q6 Q7 Q8 Q9 time (s) No deterioration for any query 62

63 TPC-H Throughput: Disk ehavior SS ase K Read Page reads and seeks reduced Time SS ase Seeks Per Sec 63 Database Research Group / IM T.J. Watson Research Time Center

64 Conclusions Mechanism for better cache reuse and reduced I/O to increase throughput and reduce latency for ad-hoc index scan-heavy multi-query workloads Inter-SISCAN cache locality improved via: Starting new SISCANs near similar (speed/key range) running scans Speed control of SISCANs to reduce drift SISCAN-based priority hints to bufferpool manager Fulfills requirements: Can handle dynamic heterogeneous workloads Easy integration in architecture 64

65 Thank you! Contact: Database Research Group IM T.J. Watson Research Center 65