This is a presentation that was held at ICPE 2013, Prague, 24/04/2013
Full paper and additional information available at:
http://msrg.org/papers/ICPE2013Rabl
Abstract:
The Star Schema Benchmark (SSB), has been widely used to evaluate the performance of database management systems when executing star schema queries. SSB, based on the well known industry standard benchmark TPC-H, shares some of its drawbacks, most notably, its uniform data distributions. Today’s systems rely heavily on sophisticated cost-based query optimizers to generate the most efficient query execution plans. A benchmark that evaluates optimizer’s capability to generate optimal execution plans under all circumstances must provide the rich data set details on which optimizers rely (uniform and non-uniform distributions, data sparsity, etc.). This is also true for other database system parts, such as indices and operators, and ultimately holds for an end-to-end benchmark as well. SSB’s data generator, based on TPC-H’s dbgen, is not easy to adapt to different data distributions as its meta data and actual data generation implementations are not separated. In this paper, we motivate the need for a new revision of SSB that includes non-uniform data distributions. We list what specific modifications are required to SSB to implement non-uniform data sets and we demonstrate how to implement these modifications in the Parallel Data Generator Framework to generate both the data and query sets.
Beginners Guide to TikTok for Search - Rachel Pearson - We are Tilt __ Bright...
Variations of the Star Schema Benchmark to Test the Effects of Data Skew on Query Performance
1. Variations of the Star
Schema Benchmark to Test
the Effects of Data Skew on
Query Performance
T IL M AN N R ABL , M EIKEL PO ESS, H AN S - AR N O
J AC O BSEN , PAT R IC K AN D EL IZABETH O’N EIL
ICPE 2013, PRAGUE, 24/04/2013
MIDDLEWARE SYSTEMS
RESEARCH GROUP
MSRG.ORG
2. Real Life Data is
Distributed Uniformly…
Well, Not Really
◦ Customers zip codes typically clustered around metropolitan areas
◦ Seasonal items (lawn mowers, snow shovels, …) sold mostly during specific
periods
◦ US retail sales:
◦ peak during Holiday Season
◦ December sales are 2x of
January sales
Source: US Census Data
RABL, POESS, JACOBSEN, O'NEIL, O'NEIL - SSB SKEW VARIATIONS
2
4. How Can Skew Effect
Database Systems?
Data placement
◦ Partitioning
◦ Indexing
Data structures
◦ Tree balance
◦ Bucket fill ratio
◦ Histograms
Optimizer finding the optimal query plan
◦ Index vs. non-index driven plans
◦ Hash join vs. merge join
◦ Hash group by vs. sort group by
RABL, POESS, JACOBSEN, O'NEIL, O'NEIL - SSB SKEW VARIATIONS
4
5. Agenda
Data Skew in Current Benchmarks
Star Schema Benchmark (SSB)
Parallel Data Generation Framework (PDGF)
Introducing Skew in SSB
RABL, POESS, JACOBSEN, O'NEIL, O'NEIL - SSB SKEW VARIATIONS
5
6. Data Skew in Benchmarks
TPC-D (1994-1999): only uniform data
◦ SIGMOD 1997 - “Successor of TPC-D
should include data skew”
◦ No effect until …
TPC-DS (released 2012)
◦ Contains comparability zones
◦ Not fully utilized
TPC-D/H variations
◦ Chaudhuri and Narayasa: Zipfian distribution on all columns
◦ Crolotte and Ghazal: comparability zones
Still lots of open potential
RABL, POESS, JACOBSEN, O'NEIL, O'NEIL - SSB SKEW VARIATIONS
6
7. Star Schema Benchmark I
Star schema version of TPC-H
◦
◦
◦
◦
Merged Order and Lineitem
Date dimension
Dropped Partsupp
Selectivity hierarchies
◦ C_City C_Nation C_Region
◦ …
RABL, POESS, JACOBSEN, O'NEIL, O'NEIL - SSB SKEW VARIATIONS
7
8. Star Schema Benchmark II
Completely new set of queries
4 flights of 3-4 queries
◦ Designed for functional coverage and selectivity coverage
◦ Drill down in dimension hierarchies
◦ Predefined selectivity
Q1.1
select
from
where
and
and
and
sum(lo_extendedprice*lo_discount) as revenue
lineorder, date
lo_orderdate = d_datekey
d_year = 1993
lo_discount between 1 and 3
lo_quantity < 25;
Q1.2
select
from
where
and
and
and
sum(lo_extendedprice*lo_discount) as revenue
lineorder, date
lo_orderdate = d_datekey
d_yearmonthnum = 199301
lo_discount between 1 and 3
lo_quantity between 26 and 35;
RABL, POESS, JACOBSEN, O'NEIL, O'NEIL - SSB SKEW VARIATIONS
8
9. Parallel Data Generation
Framework
Generic data generation framework
Relational model
◦ Schema specified in configuration file
◦ Post-processing stage for alternative representations
Repeatable computation
◦ Based on XORSHIFT random number generators
◦ Hierarchical seeding strategy
Frank, Poess, and Rabl: Efficient Update Data Generation for DBMS Benchmarks. ICPE '12.
Rabl and Poess: Parallel Data Generation for Performance Analysis of Large, Complex RDBMS. DBTest '11.
Poess, Rabl, Frank, and Danisch: A PDGF Implementation for TPC-H. TPCTC '11.
Rabl, Frank, Sergieh, and Kosch: A Data Generator for Cloud-Scale Benchmarking. TPCTC '10.
RABL, POESS, JACOBSEN, O'NEIL, O'NEIL - SSB SKEW VARIATIONS
9
11. Opportunities to Inject Data
Skew in
Foreign key relations
◦ E.g., L_PARTKEY
One fact table measures
◦ E.g., L_Quantity
Single dimension hierarchy
◦ E.g., P_Brand → P_Category → P_Mfgr
Multiple dimension hierarchies
◦ E.g., City → Nation in Supplier and Customer
Experimental methodology
◦
◦
◦
◦
One experiment series for each of the above
Comparison to original SSB
Comparison of index-forced, non-index, and automatic optimizer mode
SSB scale factor 100 (100 GB), x86 server
RABL, POESS, JACOBSEN, O'NEIL, O'NEIL - SSB SKEW VARIATIONS
11
12. Skew in Foreign Key
Relations
Very realistic
Easy to implement in PDGF
◦ Just add a distribution to the reference
<distribution name="Exponential“ lambda="0.26235" />
But!
Dimension attributes uniformly distributed
Dimension keys uncorrelated to dimension attributes
Very limited effect on selectivity
Focus on attributes in selectivity predicates
RABL, POESS, JACOBSEN, O'NEIL, O'NEIL - SSB SKEW VARIATIONS
12
13. Skew in Fact Table Measure
– Lo_Quantity
Lo_Quantity distribution
◦ Values range between 0 and 50
◦ Originally uniform distribution with:
◦ P(X=x)=0.02
◦ Coefficient of variation of 0.00000557
◦ Proposed skewed distribution with:
◦
P(X
x)
0 .3
1 .3
x
Query 1.1
◦ lo_quantity < x, x ∈ [2, 51]
Results
◦ Switches too early to non-index plan
◦ Switches too late to non-index plan
◦ Optimizer agnostic to distribution
RABL, POESS, JACOBSEN, O'NEIL, O'NEIL - SSB SKEW VARIATIONS
13
14. Skew in Single Dimension
Hierarchy - Part
P_Category distribution
◦ Uniform P(X=x)=0.04
◦ Skewed P(X=x)= 0.01 - 48.36
◦ Probabilities explicitly defined
Query 2.1
◦ Restrictions on two dimensions
Results uniform case
◦ Index driven superior
◦ Optimizer chooses non-index driven
Results skewed case
◦ Switches too early to non-index
plan
RABL, POESS, JACOBSEN, O'NEIL, O'NEIL - SSB SKEW VARIATIONS
14
15. Skew in Multiple Dimension
Hierarchies – S_City &
C_City
Skewed S_City & C_City
◦ Probabilites exponentially
distributed
Query 3.3
Join Cardinality
Elapsed Time
◦ Restrictions on 3 dimensions
◦ Variation on Supplier and Customer
city
Results uniform and skewed cases
◦ Automatic plan performs best
◦ Cross over between automatic
uniform and skewed too late
RABL, POESS, JACOBSEN, O'NEIL, O'NEIL - SSB SKEW VARIATIONS
15
16. Conclusion & Future Work
PDGF implementation of SSB
Introduction of skew in SSB
Extensive performance analysis
◦ Several interesting optimizer effects
◦ Performance impact of skew
Future Work
Further analysis on impact of skew
Skew in query generation
Complete suite for testing skew effects
RABL, POESS, JACOBSEN, O'NEIL, O'NEIL - SSB SKEW VARIATIONS
16
17. Thanks
Questions?
Download and try PDGF:
http://www.paralleldatageneration.org
(scripts used in the study available on website above)
RABL, POESS, JACOBSEN, O'NEIL, O'NEIL - SSB SKEW VARIATIONS
17