[PPT] - SIMD Vectorized Hashing for Grouped Aggregation Bala Gurumurthy, PowerPoint Presentation

SLIDE 1

1 16.05.2017 OVGU Präsentation

SIMD Vectorized Hashing for Grouped Aggregation

1 Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SLIDE 2

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Grouped Aggregation

2

Commonly-used and time-consuming operation

Based on analysis by Boncz et al.[1]

SLIDE 3

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Grouped Aggregation

2

Commonly-used and time-consuming operation
All input must be consumed for single output

Based on analysis by Boncz et al.[1]

SLIDE 4

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Grouped Aggregation

2

Commonly-used and time-consuming operation
All input must be consumed for single output
Faster input processing = higher throughput

Based on analysis by Boncz et al.[1]

SLIDE 5

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Grouped Aggregation

2

Commonly-used and time-consuming operation
All input must be consumed for single output
Faster input processing = higher throughput
Improving underlying technique improves efficiency

Based on analysis by Boncz et al.[1]

SLIDE 6

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Capability in Modern Processors

3

SIMD – Single Instruction Multiple Data

SLIDE 7

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Capability in Modern Processors

SIMD – Single Instruction Multiple Data
Allows vectorized execution in modern processors
Reduces overall execution time of an operation

3

SLIDE 8

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Capability in Modern Processors

SIMD – Single Instruction Multiple Data
Allows vectorized execution in modern processors
Reduces overall execution time of an operation
SIMD is shown to increase throughput in orders of magnitude for

DBMS operation [2] [3]

SIMD accelerated selection [3] 3

SLIDE 9

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD for Grouped Aggregation

SIMD acceleration of hashing techniques improves throughput

4 + Group-By = High throughput

SLIDE 10

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD for Grouped Aggregation

SIMD acceleration of hashing techniques improves throughput

How to incorporate SIMD for Grouped Aggregation? What is the impact of SIMD? + Group-By = High throughput 4

SLIDE 11

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hash Based Aggregation

Grouped aggregation commonly implemented using hashing

techniques

5

SLIDE 12

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hash Based Aggregation

Grouped aggregation commonly implemented using hashing

techniques

5

SLIDE 13

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hash Based Aggregation

Grouped aggregation commonly implemented using hashing

techniques

Separates groups into buckets

5

SLIDE 14

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hash Based Aggregation

Grouped aggregation commonly implemented using hashing

techniques

Separates groups into buckets
Aggregation done within each buckets

5

SLIDE 15

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hash Based Aggregation: Example

Key Aggregate

h(x)

Hash function

6

SLIDE 16

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hash Based Aggregation: Example

Input: 3 h(x) 6

SLIDE 17

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hash Based Aggregation: Example

Input: 3 h(x)

3

6

SLIDE 18

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hash Based Aggregation: Example

Input: 3 h(x) 6

SLIDE 19

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hash Based Aggregation: Example

Input: 3 h(x)

3

6

SLIDE 20

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hash Based Aggregation: Example

Input: 3 h(x) 6

SLIDE 21

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Collision: Increasing Complexity

Not all keys have unique location

h(x) 7

SLIDE 22

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Collision: Increasing Complexity

Not all keys have unique location
Two keys might hash to same slots

h(x)

1 11

7

SLIDE 23

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Collision: Increasing Complexity

Not all keys have unique location
Two keys might hash to same slots

h(x)

11

7

SLIDE 24

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Collision: Increasing Complexity

Not all keys have unique location
Two keys might hash to same slots
Hash table must be probed for alternative location

h(x)

11

7

SLIDE 25

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Collision: Increasing Complexity

Not all keys have unique location
Two keys might hash to same slots
Hash table must be probed for alternative location

h(x)

11

# of probes : 4 7

SLIDE 26

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Collision: Increasing Complexity

Not all keys have unique location
Two keys might hash to same slots
Hash table must be probed for alternative location

h(x) Probing is time consuming 7

SLIDE 27

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD for Probing

Multiple slots are probed in an instant using SIMD

h(x) 7

SLIDE 28

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD for Probing

h(x) # of probes : 1

Multiple slots are probed in an instant using SIMD
Reduces overall number of probes

7

SLIDE 29

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Accelerated Hash Probing

Each hashing techniques have their own collision resolution mechanism

8

SLIDE 30

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Accelerated Hash Probing

Each hashing techniques have their own collision resolution mechanism
We use open-addressing hashing techniques

8

SLIDE 31

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Accelerated Hash Probing

Each hashing techniques have their own collision resolution mechanism
We use open-addressing hashing techniques
Have constant hashtable size

8

SLIDE 32

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Accelerated Hash Probing

Each hashing techniques have their own collision resolution mechanism
We use open-addressing hashing techniques
Have constant hashtable size
Suitable for SIMD

8

SLIDE 33

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Accelerated Hash Probing

Each hashing techniques have their own collision resolution mechanism
We use open-addressing hashing techniques
Have constant hashtable size
Suitable for SIMD

Hashing techniques used are ➢ Cuckoo hashing ➢ Linear probing ➢ Two-choice hashing ➢ Hopscotch hashing 8

SLIDE 34

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Cuckoo Hashing

Stores keys in multiple hash tables

9

SLIDE 35

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Cuckoo Hashing

Stores keys in multiple hash tables
On collision swaps current value to alternative tables

Input : 1 9

SLIDE 36

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Cuckoo Hashing

Stores keys in multiple hash tables
On collision swaps current value to alternative tables
Might form swap loop; solved using a threshold

Input : 111 9

SLIDE 37

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Cuckoo Hashing

Stores keys in multiple hash tables
On collision swaps current value to alternative tables
Might form swap loop; solved using a threshold
Has constant look-up time

Input : 1 9

SLIDE 38

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Grouped Aggregation using Cuckoo Hashing

Probe for key in each hash table

9

SLIDE 39

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Grouped Aggregation using Cuckoo Hashing

Probe for key in each hash table
If found update aggregate

9

SLIDE 40

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Grouped Aggregation using Cuckoo Hashing

Probe for key in each hash table
If found update aggregate
Else, insert the key

9

SLIDE 41

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Optimization of Cuckoo Hashing

Hash function computed for multiple tables

9

SLIDE 42

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Optimization of Cuckoo Hashing

Hash function computed for multiple tables
Multiple slots probed in parallel

9

SLIDE 43

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Linear Probing

Straight forward approach for collision resolution

10

SLIDE 44

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Linear Probing

Straight forward approach for collision resolution
Probes hash table linearly for alternative location

10

SLIDE 45

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Linear Probing

Straight forward approach for collision resolution
Probes hash table linearly for alternative location
Empty location encountered : insert key

10

SLIDE 46

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Optimization of Linear Probing

Multiple slot is probed

10

SLIDE 47

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Optimization of Linear Probing

Multiple slot is probed
Comparison mask used for updating aggregate

10

SLIDE 48

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Two-choice Hashing

11 h1(x)

11

h1(x)

Improvement on linear probing

SLIDE 49

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Two-choice Hashing

h1(x)

11

h1(x)

Improvement on linear probing
Two hashing function for same hash table (can be more)

11

SLIDE 50

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Two-choice Hashing

Improvement on linear probing
Two hashing function for same hash table (can be more)
If all slots occupied, probing done from both slots

h1(x)

11

h1(x) 11

SLIDE 51

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Optimization of Two-Choice Hashing

SIMD hash function from cuckoo hashing

11

SLIDE 52

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Optimization of Two-Choice Hashing

SIMD hash function from cuckoo hashing
SIMD probing from linear probing

11

SLIDE 53

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hopscotch Hashing

Has a limited probe length known as neighborhood

Neighborhood : 3

x y z 12

SLIDE 54

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hopscotch Hashing

Has a limited probe length known as neighborhood
A key will be available within neighborhood distance from hash slot

Neighborhood : 3 key

x y z 12

SLIDE 55

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hopscotch Hashing

Has a limited probe length known as neighborhood
A key will be available within neighborhood distance from hash slot
If no slot available within neighborhood, table is rearranged by swapping keys

Neighborhood : 3

x (0) y (1) z (2)

Input : New key Key (0)

12

SLIDE 56

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hopscotch Hashing

Has a limited probe length known as neighborhood
A key will be available within neighborhood distance from hash slot
If no slot available within neighborhood, table is rearranged by swapping keys

Neighborhood : 3

x (0) y (1) z (2)

Input : New key Key (0)

12

SLIDE 57

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hopscotch Hashing

Has a limited probe length known as neighborhood
A key will be available within neighborhood distance from hash slot
If no slot available within neighborhood, table is rearranged by swapping keys

Neighborhood : 3

x (0) y (1)

Input : New key

z (2)

Key (0)

12

SLIDE 58

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hopscotch Hashing

Has a limited probe length known as neighborhood
A key will be available within neighborhood distance from hash slot
If no slot available within neighborhood, table is rearranged by swapping keys

Neighborhood : 3 key New key

x y z 12

SLIDE 59

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Optimization of Hopscotch Hashing

SIMD for Probing
Uses SIMD linear probing

13

SLIDE 60

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Optimization of Hopscotch Hashing

SIMD for Probing
Uses SIMD linear probing
SIMD for Insertion
Swapping multiple values in an instant

13

SLIDE 61

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Optimization of Hopscotch Hashing

SIMD for Probing
Uses SIMD linear probing
SIMD for Insertion
Swapping multiple values in an instant

13

SLIDE 62

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Optimization of Hopscotch Hashing

SIMD for Probing
Uses SIMD linear probing
SIMD for Insertion
Swapping multiple values in an instant
Collect values using SIMD Gather

13

SLIDE 63

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Optimization of Hopscotch Hashing

SIMD for Probing
Uses SIMD linear probing
SIMD for Insertion
Swapping multiple values in an instant
Collect values using SIMD Gather
Swap them

13

SLIDE 64

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Optimization of Hopscotch Hashing

SIMD for Probing
Uses SIMD linear probing
SIMD for Insertion
Swapping multiple values in an instant
Collect values using SIMD Gather
Swap them
Store back

13

SLIDE 65

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Evaluation Setup

Environment
Processor : Intel Octa Core Xeon E5-2630
OS: Linux
SIMD : AVX2 instruction set
Grouped Aggregation
Keys : 32 bit integers (0 is not valid)
Aggregation : count()
Hashing technique
Hash function : Multiplicative function – h(x) = Ax%tableSize
A => knuth’s number
Data Distribution
Uniform random
Sequential
Unique random
Moving cluster

14

SLIDE 66

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Unique Random Distribution

Serial linear probing has worse time
Serial hopscotch hashing is efficient
Vectorized two-choice hashing

competes with Hopscotch hashing

14

SLIDE 67

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Uniform Random Distribution

Both serial and vectorized hopscotch

has worse efficiency

Vectorized with the worst time
Vectorized Two-choice hashing has

best execution time

15

SLIDE 68

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Uniform Random Distribution

Both serial and vectorized hopscotch

has worse efficiency

Vectorized with the worst time
Vectorized Two-choice hashing has

best execution time

Other distributions have the same characteristics 15

SLIDE 69

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Overall Speed-up

SIMD has worse impact on hopsctoch

hashing

Nearly 2x slower
Due to random access for

insertion

Linear probing has highest SIMD

impact

16

SLIDE 70

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Summary

Hash probing reduces efficiency of grouped aggregation

17

SLIDE 71

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Summary

Hash probing reduces efficiency of grouped aggregation
SIMD accelerated probing not always improves efficiency
Specifically, hopscotch hashing has negative impact
Linear probing has up to 3.5x speed up due to SIMD

17

SLIDE 72

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Summary

Hash probing reduces efficiency of grouped aggregation
SIMD accelerated probing not always improves efficiency
Specifically, hopscotch hashing has negative impact
Linear probing has up to 3.5x speed up due to SIMD
Hashing technique related parameters also improve the efficiency

17

SLIDE 73

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Summary

Hash probing reduces efficiency of grouped aggregation
SIMD accelerated probing not always improves efficiency
Specifically, hopscotch hashing has negative impact
Linear probing has up to 3.5x speed up due to SIMD
Hashing technique related parameters also improve the efficiency
Further, improvement can be extended using SIMD for multiple

insertion

17

SLIDE 74

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Questions? Thank You

SLIDE 75

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

References

[1] Boncz, P., Neumann, T., & Erling, O. (2014). TPC-H Analyzed: Hidden Messages and Lessons Learned from an Influential Benchmark. Springer International Publishing. [2] Orestis Polychroniou, Arun Raghavan, and Kenneth A. Ross. (2015). Rethinking SIMD Vectorization for In- Memory Databases. In Proceedings of the 2015 ACM SIGMOD International Conference on Management of Data (SIGMOD '15) [3] Broneske, D., Meister, A., & Saake, G. (2017). Hardware-Sensitive Scan Operator Variants for Compiled Selection Pipelines. BTW

SLIDE 76

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Cuckoo Hashing – Table Structure

Table Structure

–

Packed key and payload

–

Each bucket contain one key, payload pack

Pack size = SIMD vector size
Ross et al. explored SIMD probing [4]

SLIDE 77

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Adapted from [6] 7

SIMD Acclerated Cuckoo Hashing

Hash Function Compute (Multiplicative hashing) Table Probe

SLIDE 78

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Search key – K is duplicated

Adapted from [6] 7

SIMD Acclerated Cuckoo Hashing

SLIDE 79

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Hashing function
Returns two

bucket positions

Adapted from [6] 7

SIMD Acclerated Cuckoo Hashing

SLIDE 80

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Acclerated Cuckoo Hashing

Probe values in

the table

Compare with

search key

Adapted from [6] 7

SLIDE 81

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Acclerated Cuckoo Hashing

Comparison result(MASK) are added to payload in the slots

Adapted from [6] 7

SLIDE 82

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Acclerated Cuckoo Hashing

If result of all the masks are 0, insertion is performed

Adapted from [6] 7

SLIDE 83

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Linear Probing

Table Structure

–

Has SoA (Structure of Array format)

–

Key and payload in individual array

–

payload in corresponding key position 8

SLIDE 84

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Linear Probing – SIMD Code Optimization

9

SLIDE 85

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Linear Probing – SIMD Code Optimization

Search key – K is duplicated

9

SLIDE 86

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Linear Probing – SIMD Code Optimization

Scalar hash function
Table bucket is selected

9

SLIDE 87

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Linear Probing – SIMD Code Optimization

Compare slot values with search key

9

SLIDE 88

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Linear Probing – SIMD Code Optimization

Add comparison results with payloads

9

SLIDE 89

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

Linear Probing – SIMD Code Optimization

Comparison result Equality – return Inequality – Search next slot Empty location – Insert the value

9

SLIDE 90

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Accelerated Hopscotch Hashing

SIMD for Probing

–

Uses SIMD linear probing

–

Probe key within Neighborhood

–

Probe empty space outside

SIMD For Insertion

–

Starts when empty space is found

SLIDE 91

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Accelerated Hopscotch Hashing

SIMD for Probing

–

Uses SIMD linear probing

–

Probe key within Neighborhood

–

Probe empty space outside

SIMD For Insertion

–

Starts when empty space is found

–

Swap previous values until empty space is inside neighborhood

SLIDE 92

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Accelerated Hopscotch Hashing

SIMD for Probing

–

Uses SIMD linear probing

–

Probe key within Neighborhood

–

Probe empty space outside

SIMD For Insertion

–

Starts when empty space is found

–

Swap previous values until empty space is inside neighborhood

Swap array holds key to swap Use SIMD gather to collect keys from hash table

SLIDE 93

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Accelerated Hopscotch Hashing

SIMD for Probing

–

Uses SIMD linear probing

–

Probe key within Neighborhood

–

Probe empty space outside

SIMD For Insertion

–

Starts when empty space is found

–

Swap previous values until empty space is inside neighborhood

Using SIMD shift to move the keys one step

SLIDE 94

04.09.2018

SIMD vectorized Hashing for Grouped Aggregation

Bala Gurumurthy, David Broneske, Marcus Pinnecke, Gabriel Campero Durand and Gunter Saake

SIMD Accelerated Hopscotch Hashing

SIMD for Probing

–

Uses SIMD linear probing

–

Probe key within Neighborhood

–

Probe empty space outside

SIMD For Insertion

–

Starts when empty space is found

–

Swap previous values until empty space is inside neighborhood

Using the position the new values are written back