This is a presentation that was held at the Second TPC Technology Conference on Performance Evaluation & Benchmarking, 17.09.2010, Singapore. Abstract: In many fields of research and business data sizes are breaking the petabyte barrier. This imposes new problems and research possibilities for the database community. Usually, data of this size is stored in large clusters or clouds. Although clouds have become very popular in recent years, there is only little work on benchmarking cloud applications. We present a data generator for cloud sized applications. Its architecture makes the data generator easy to extend and to configure. A key feature is the high degree of parallelism that allows linear scaling for arbitrary numbers of nodes.We show how distributions, relationships and dependencies in data can be computed in parallel with linear speed up.

1. T. Rabl, M. Frank, H. Mousselly Sergieh, and H. Kosch
17.09.2010
Second TPC Technology Conference on Performance Evaluation &
Benchmarking

2. 
Data sizes grow beyond petabyte barrier
 Cloud computing to the rescue

Need for cloud-scale benchmarking
 Need for realistic, cloud-scale data sets

Problems:
 Generating petabytes
 Storing petabytes
 Transporting petabytes

Solution:
 Parallel, on-site generation
A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
2

3. 
3 tables with primary and foreign keys
Non-uniform distributions (lognormal)
Replicated data

How to generate consistent data in parallel?


A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
3

4. 
No references
 Only uncorrelated data
 Simple statistical references

Scanned references
 Read data to generate reference
 Generate tuple pairs

Computed references
 Data is generated deterministically
 Compute data to generate reference
Parallel Data Generation Framework
A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
4

5. 
Parallel pseudo random number generator
 Deterministic
 Fast skip ahead

Generation of values as a function
 Deterministic

Seeds + row id + generator allows
recalculation
 Every value can be computed independently
A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
5

6. seed
t_id
Table RNG
seed
c_id
Column RNG
seed
r_id
Row RNG
rn



Generator
user
row_id user_id degree_program …
1
2
3
Seeding strategy
All seeds can be cached
Generation of value in row n with n-th
random number
A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
6

7. seed
id
Table RNG
seed
id
Column RNG
seed
r_id
Lognormal
rn
seed
rn



Row RNG
r_id
seminar_user
row_id user_id degree_program …
1
2
3
Row RNG
Generator
Reference generation
Lognormal indexes row_id of user
Equal seeds and lognormal generator for user_id
A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
7

8. 

Java based
Plug-in concept
 Generators
 Distributions
 RNGs
 Output
 Scheduler

TPC-H plug-in
A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
8

9. 

XML file
Reflects SQL schema
 Tables
 Fields




Seed
Size
Scale factor
Output
<project name="simpleUserSeminar">
[..]
<table name=“seminar_user">
<size>201754</size>
<fields>
<field name="user_id">
<type>java.sql.Types.INTEGER</type>
<reference>
<referencedField>user_id</referencedField>
<referencedTable>user</referencedTable>
</reference>
<generator name="DefaultReferenceGenerator">
<distribution name="LogNormal">
<mu>7.60021</mu><sigma>1.40058</sigma>
</distribution>
</generator>
</field>
<field name="degree_program">
[..]
A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
9

10. 


Every value can be computed independently
<?xml version="1.0" encoding="UTF-8"?>
No communication
<nodeConfig>
<nodeNumber>1</nodeNumber>
Workload distribution
<nodeCount>2</nodeCount>
 Configuration for every node
<workers>2</workers>
</nodeConfig>
 Automatic distribution
seminar_user
Cluster/Cloud
Nodes
Processors
100000 rows
Row 1 – 50000
Row 1 –
25000
Row 25001 –
50000
Row 50001 – 100000
Row 50001 –
75000
A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
Row 75001 –
100000
10

11. 
16 node HPC cluster
 2 Intel Xeon QuadCore processors
 16GB RAM
 2 x 74GB HDD, RAID 0

SetQuery data set
 1 table “Bench”, 21 columns
 12 Random numbers, 8 strings

TPC-H data set
 8 tables, 61 columns
 Data types: integer, char, varchar, decimal, date
A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
11

12. 


SetQuery data set
100 GB data set (SF 460)
1 to 16 nodes
A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
12

13. 
TPC-H data set
 Data sizes: 1 GB, 10 GB, 100 GB

Single node (8 cores)
A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
13

14. 


Parallel data generation framework
Linear speed up
Fast generation
 Large data sets
 Realistic data


Independent generation of tables / colums /
values
Easy configuration and extension
A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
14

15. 
More implementation





generators, distributions
Benchmarks
Graphical user interface
Scheduler
Query generator
 Consistent inserts, updates, deletes
 Precomputed query results
 Time series
A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
15

17. 
SetQuery data set
 Data sizes: 220 MB, 2.2 GB, 10 GB, 22 GB, 100 GB

Single node
A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
17

18. 
TPC-H data set
 Data sizes: 1 GB, 10 GB, 100 GB, 1TB

1, 10, 16 nodes
A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
18

19. A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
19

20. A Data Generator for Cloud-Scale Benchmarking - Tilmann Rabl
20