Parallel DBs. April 25, 2017

Transcription

1 Parallel DBs April 25,

2 Why Scale Up? Scan of 1 PB at 300MB/s (SATA r2 Limit) (x1000) ~1 Hour ~3.5 Seconds 2

3 Data Parallelism Replication Partitioning A A A A B C 3

4 Operator Parallelism Pipeline Parallelism: A task breaks down into stages; each machine processes one stage. Sequential Operation Sequential Operation Sequential Operation Partition Parallelism: Many machines doing the same thing to different pieces of data. Sequential Operation Sequential Operation Sequential Operation 4

5 Types of Parallelism Both types of parallelism are natural in a database management system. SELECT SUM( ) FROM Table WHERE Sequential Operation Sequential Operation Sequential Operation Sequential Operation Sequential Operation Sequential Operation Sequential Operation Sequential Operation Sequential Operation Sequential Operation LOAD SELECT AGG Combine 5

6 DBMSes: The First Success Story Every major DBMS vendor has a version. Reasons for success: Bulk Processing (Partition -ism). Natural Pipelining in RA plan. Users don t need to think in. 6

7 Types of Speedup Speed-up -ism More resources = proportionally less time spent. Scale-up -ism More resources = proportionally more data processed. Response Time Throughput # of Nodes # of Nodes 7

8 Parallelism Models CPU Memory Disk 8

9 Parallelism Models CPU Memory Disk How do the nodes communicate? 9

10 Parallelism Models Option 1: Shared Memory available to all CPUs CPU Memory Disk e.g., a Multi-Core/Multi-CPU System 10

11 Parallelism Models Option 2: Non-Uniform Memory Access. CPU Memory Disk Used by most AMD servers 11

12 Parallelism Models Option 3: Shared Disk available to all CPUs CPU Memory Disk Each node interacts with a disk on the network. 12

13 Parallelism Models Option 4: Shared Nothing in which all communication is explicit. CPU Memory Disk Examples include MPP, Map/Reduce. Often used as basis for other abstractions. 13

14 Parallelizing OLAP - Parallel Queries OLTP - Parallel Updates 14

15 Parallelizing OLAP - Parallel Queries OLTP - Parallel Updates 15

16 Parallelism & Distribution Distribute the Data Redundancy Faster access Parallelize the Computation Scale up (compute faster) Scale out (bigger data) 16

17 Operator Parallelism General Concept: Break task into individual units of computation. Challenge: How much data does each unit of computation need? Challenge: How much data transfer is needed to allow the unit of computation? Same challenges arise in Multicore, CUDA programming. 17

18 Parallel Data Flow A No Parallelism 18

19 Parallel Data Flow A 1 AN N-Way Parallelism 19

20 Parallel Data Flow B 1 BN??? A 1 AN Chaining Parallel Operators 20

21 Parallel Data Flow B 1 BN A 1 AN One-to-One Data Flow ( Map ) 21

22 Parallel Data Flow B 1 BN A 1 AN One-to-One Data Flow 22

23 Parallel Data Flow Extreme 1 All-to-All All nodes send all records to all downstream nodes B A 1 BN 1 AN Many-to-Many Data Flow 23 Extreme 2 Partition Each record goes to exactly one downstream node

24 Parallel Data Flow B A 1 AN Many-to-One Data Flow ( Reduce/Fold ) 24

25 Parallel Operators Select Project Union (bag) What is a logical unit of computation? (1 tuple) Is there a data dependency between units? (no) 25

26 Parallel Operators Select Project Union (bag) A 1 AN 1/N Tuples 1/N Tuples 26

27 Parallel Joins FOR i IN 1 to N FOR j IN 1 to K Partition JOIN(Block i of R, Partition Block j of S) One Unit of Computation 27

28 Parallel Joins K Partitions of S N Partitions of R Block 1 of R Block 1 of S N Block N of R Block 1 of S K K Block 1 of R Block K of S N Block N of R Block K of S 28

29 Practical Concerns UNION R1 S1 R1 S2 R2 S1 RN SM R1 R2 RN S1 S2 SM Where does the computation happen? How does the data get there? 29

30 Distributing the Work Let s start simple what can we do with no partitions? B R S R and S may be any RA expression 30

31 Distributing the Work B R S Node 1 No Parallelism! 31

32 Distributing the Work All of R and All of S get sent! B Node 3 R Node 1 S Node 2 Lots of Data Transfer! 32

33 Distributing the Work All of R get sent B R Node 1 S Node 2 Better! We can guess whether R or S is smaller. 33

34 Distributing the Work What can we do if R is partitioned? U B B R1 S R2 34

35 Distributing the Work There are lots of partitioning strategies, but this one is interesting. U B B R1 Node 1 S R2 Node 2 Node 3 35

36 Distributing the Work it can be used as a model for partitioning S U B B R1 Node 1 S1 R2 Node 2 Node 3 36

37 Distributing the Work it can be used as a model for partitioning S U B B R1 Node 1 S2 R2 Node 2 Node 3 37

38 Distributing the Work and neatly captures the data transfer issue. U B B R1 Node 1 S R2 Node 2 Node 3 38

39 Hash(R.B)%4 0 Hash(S.B)%4 Parallel Joins X X X R B S: Which Partitions of S Join w/ Bucket 0 of R? 39

40 Parallel Joins R.B S.B B<25 B<25 25 B<50 50 B<75 75 B 25 B<50 X 50 B<75 X X 75 B X R R.B < S.B S: X Which Partitions of S Can Produce Output? 40 X

41 Can we further reduce the amount of data sent? 41

42 Sending Hints Rk B Si The naive approach Node 1 Node 2 Rk Si 42

43 Sending Hints Rk B Si The naive approach Node 1 Send me Node 2 Rk Rk Si 43

44 Sending Hints Rk B Si The naive approach Rk Node 1 Node 2 Rk Si 44

45 Sending Hints Rk B Si The smarter approach πb( ) Si Node 1 Node 2 Rk Si 45

46 Sending Hints Rk B Si The smarter approach πb( ) Si Node 1 Node 2 πb( ) Rk Rk Si Si 46

47 Sending Hints Rk B Si The smarter approach Node 1 Node 2 <1,A> <2,B> <2,C> <3,D> <4,E> 47 <2,X> <3,Y> <6,Y>

48 Sending Hints Rk B Si The smarter approach Node 1 Node 2 <1,A> <2,B> <2,C> <3,D> <4,E> 48 Send me rows with a B of 2,3, or 6 <2,X> <3,Y> <6,Y>

49 Sending Hints Rk B Si The smarter approach Node 1 Node 2 <1,A> <2,B> <2,C> <3,D> <4,E> <2,B> <2,C> <3,D> 49 Send me rows with a B of 2,3, or 6 This is called a semi-join. <2,X> <3,Y> <6,Y>

50 Sending Hints Now Node 1 sends as little data as possible but Node 2 needs to send a lot of data. Can we do better? 50

51 Sending Hints Rk B Si Strategy 1: Parity Bits Node 1 Node 2 <1,A> <2,B> <2,C> <3,D> <4,E> <2,X> 0 <6,Y>

52 Sending Hints Rk B Si Strategy 1: Parity Bits Node 1 Node 2 <1,A> <2,B> <2,C> <3,D> <4,E> Send me data with a parity bit of 0 0 <2,X> 0 <6,Y>

53 Sending Hints Rk B Si Strategy 1: Parity Bit Node 1 sending too much is ok! (Node 2 still needs to compute B) Node 1 Node 2 <1,A> <2,B> <2,C> <3,D> <4,E> <2,B> <2,C> <4,E> 53 Send me data with a parity bit of 0 Problem: One parity bit is too little 0 <2,X> 0 <6,Y>

54 Sending Hints Rk B Si Strategy 1: Parity Bit Node 1 Node 2 <1,A> <2,B> <2,C> <3,D> <4,E> Problem: One parity bit is too little 54 0 <2,X> 1 <3,Y> 0 <6,Y>

55 Sending Hints Rk B Si Strategy 2: Parity Bits Node 1 Node 2 <1,A> <2,B> <2,C> <3,D> <4,E> <2,B> <2,C> <3,D> 55 Send me data with parity bits 10 or 11 Problem: Almost as much data as πb 10<2,X> 11<3,Y> 10<6,Y>

56 Sending Hints Can we summarize the parity bits? 56

57 Bloom Filters Alice Bob Carol Dave 57

58 Bloom Filters Alice Bob Carol Dave Bloom Filter 58

59 Bloom Filters Yes Is Alice part of the set? Alice Bob Carol Bloom Filter No Is Eve part of the set? Dave Bloom Filter Guarantee Test definitely returns Yes if the element is in the set Test usually returns No if the element is not in the set Yes Is Fred part of the set? 59

60 Bloom Filters A Bloom Filter is a bit vector M - # of bits in the bit vector K - # of hash functions For ONE key (or record): For i between 0 and K: bitvector[ hashi (key) % M ] = 1 Each bit vector has ~K bits set 60

61 Bloom Filters Key 1 Key 2 Key Filters are combined by Bitwise-OR e.g. (Key 1 Key 2) = How do we test for inclusion? (Key & Filter) == Key? Key (Key 1 & S) = (Key 3 & S) = X 61 (Key 4 & S) = False Positive

62 Sending Hints Rk B Si Strategy 3: Bloom Filters Node 1 Node 2 <1,A> <2,B> <2,C> <3,D> <4,E> 62 <2,X> <3,Y> <6,Y>

63 Sending Hints Rk B Si Strategy 3: Bloom Filters Node 1 with a B in the Node 2 <1,A> <2,B> <2,C> <3,D> <4,E> 63 Send me rows bloom filter summarizing the set {2,3,6} <2,X> <3,Y> <6,Y>

64 Sending Hints Rk B Si Strategy 3: Bloom Filters Node 1 <2,B> with a B in the Node 2 <1,A> <2,B> <2,C> <3,D> <4,E> <2,C> <3,D> <4,E> This is called a bloom-join. 64 Send me rows bloom filter summarizing the set {2,3,6} <2,X> <3,Y> <6,Y>

65 Parallel Aggregates Algebraic: Bounded-size intermediate state (Sum, Count, Avg, Min, Max) Holistic: Unbounded-size intermediate state (Median, Mode/Top-K Count, Count-Distinct; Not Distribution-Friendly) 65

66 Fan-In Aggregation SUM B A 1 AN 66

67 Fan-In Aggregation SUM 8 Messages A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 67

68 Fan-In Aggregation SUM 4 Messages 2 Messages (each) SUM 1 SUM 2 SUM 3 SUM 4 A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 68

69 Fan-In Aggregation SUM 2 Messages 2 Messages (each) SUM 1 SUM 2 SUM 1 SUM 2 SUM 3 SUM 4 A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 69

70 Fan-In Aggregation If Each Node Performs K Units of Work (K Messages) How Many Rounds of Computation Are Needed? LogK(N) 70

71 Fan-In Aggregation Components Combine(Intermediate1,, IntermediateN) = Intermediate <SUM1, COUNT1> <SUMN, COUNTN> = <SUM1+ +SUMN, COUNT1+ +COUNTN> Compute(Intermediate) = Aggregate Compute(<SUM, COUNT>) = SUM / COUNT 71