Community Training - Lehrstuhl für Datenbanksysteme

Lehrstuhl Informatik III: Datenbanksysteme
Community Training:
Partitioning Schemes in
Good Shape for Federated
Data Grids
Tobias Scholl, Richard Kuntschke, Angelika Reiser,
Alfons Kemper
3rd IEEE International Conference
on e-Science and Grid Computing
Bangalore, India
December 10th – 13th 2007
Lehrstuhl Informatik III: Datenbanksysteme
The AstroGrid-D Project
German Astronomy
Community Grid
http://www.gac-grid.org/
 funded by the German
Ministry of Education and
Research
 part of the D-Grid

Community Training
2
Lehrstuhl Informatik III: Datenbanksysteme
The Multiwavelength Milky Way
http://adc.gsfc.nasa.gov/mw/
Community Training
3
Lehrstuhl Informatik III: Datenbanksysteme
Data-intensive e-Science Applications

Many e-science application areas



astrophysics
geosciences
climatology
Combination of various, globally distributed
data sources
 Increasing popularity (within community and
public domain): more users
 Scalability issues with current approaches

Community Training
4
Lehrstuhl Informatik III: Datenbanksysteme
Up-coming Data-intensive Applications

Data rates


Terabytes a day/night
Petabytes a year
LOFAR
 LSST
 Pan-STARRS
 LHC
LOFAR

LHC
Community Training
5
Lehrstuhl Informatik III: Datenbanksysteme
Current Sharing in Data Grids
Data autonomy
 Policies allow partners to access data
 Each institution ensures




Availability (replication)
Scalability
Various organizational structures:




Centralized
Hierarchical
Federated
Hybrid
Images from: “Data-Intensive Grid Computing” by Srikumar Venugopal
Community Training
6
Lehrstuhl Informatik III: Datenbanksysteme
Community-Driven Data Distribution
Community Training
7
Lehrstuhl Informatik III: Datenbanksysteme
The Training Phase
Extract data from the archives
 Compute partitioning schemes
 Compare different partitioning schemes




Standard Quadtrees
Median-based Quadtrees
Zones
Community Training
8
Lehrstuhl Informatik III: Datenbanksysteme
Quadtrees
Well-known index
structure
 Recursive
decomposition
 Adaptive to data
resolution

Community Training
9
Lehrstuhl Informatik III: Datenbanksysteme
Quadtree-based Schemes: Splitting Variants
Center splitting
Median heuristics
 Always bisects each
 Compute median for
dimension
each dimension
independently
 congruent subareas
 Splitting points stored  O(n) median
algorithm
or computed
 Splitting points stored
Community Training
10
Lehrstuhl Informatik III: Datenbanksysteme
The Zones Index
(J. Gray, M. Nieto-Santisteban, A. Szalay)
 Index structure for databases
 Specific spatial clustering in zones
 Optimized for



Equi-distant partitioning



points-in-region queries
self-match, cross-match queries
Declination coordinate
Zone(ra, dec) = floor((dec + 90.0) / h)
Implemented directly in SQL
Community Training
Image by J. Gray et al.
11
Lehrstuhl Informatik III: Datenbanksysteme
Evaluation Setup
2 data sets: skewed and uniform
 Size of data sample: 0.01%, 0.1%, 1%, 10%
 Number of partitions:
4, 16, 64, 256, 1024, 4096, 8192, 16384,
32768, 65536, 131072 (24 – 217)

Community Training
12
Lehrstuhl Informatik III: Datenbanksysteme
Skewed Training Data (Dskew)
Community Training
13
Lehrstuhl Informatik III: Datenbanksysteme
Comparing Partitioning Schemes
Duration
 Average data population
 Variance in partition population
 Empty partitions
 Size of the training set
 Baseline comparison

Community Training
14
Lehrstuhl Informatik III: Datenbanksysteme
Average Data Population
average data population (%)
100
90
80
70
avg(# objects in partitions)
60
50
# objects in biggest partition
40
30
center splitting (skewed data)
median splitting (skewed data)
center splitting (uniform data)
median splitting (uniform data)
20
10
0
16
64
256
1024
4096
8192
16384
32768
65536
131072
# partitions
10% training sample, Dskew
Community Training
15
Lehrstuhl Informatik III: Datenbanksysteme
Evolution of the Partitioning Scheme
4096 partitions
8192 partitions
Community Training
16384 partitions
16
Lehrstuhl Informatik III: Datenbanksysteme
Empty Leaves
8
empty partitions (%)
7
6
5
4
3
2
center splitting
median splitting
zones
1
0
16
64
256
1024
4096
8192
16384
32768
65536
131072
# partitions
10% training sample, Dskew
Community Training
17
Lehrstuhl Informatik III: Datenbanksysteme
Size of the Training Set
4096
sample ratio
35
sample ratio
center splitting
median splitting
30
1024
25
256
20
# objects in training set
64
15
# partitions to be created
16
10
4
5
1
0
131072
16
64
256
1024
4096
8192
16384
32768
65536
empty partitions (%)
16384
# partitions
0.1% training sample, Dskew
Community Training
18
Lehrstuhl Informatik III: Datenbanksysteme
Standard Quadtree vs. Median Heuristics
Community Training
19
Lehrstuhl Informatik III: Datenbanksysteme
Evaluation – Summary
Quadtrees: good adaption to data distribution
 Quadtrees: Trade-off between data load
balancing and uniformly shaped regions
 Median-based heuristics: best data load
balancing even for skewed data sets
 Zone Index: good for uniform data sets
 Training set needs to be sufficiently large in
order not to artificially create empty partitions

Community Training
20
Lehrstuhl Informatik III: Datenbanksysteme
HiSbase
Peer-to-Peer layer
assigns data
partitions to peers
 Higher flexibility
 New peers are
integrated seamlessly

Community Training
21
Lehrstuhl Informatik III: Datenbanksysteme
Query Submission (at Peer d)
0




Determine relevant
regions: [1,3]
Select coordinator:
Region 1
Send CoordinateQuerymessage to id1:
{[1,3], SQL}
Message gets routed to
Peer a.
Community Training
5
1
3
2
4
6
22
Lehrstuhl Informatik III: Datenbanksysteme
Query Coordination (at Peer a)
Peer a is coordinator
 Contact relevant
regions
 Collect intermediate
results
 Send complete result
to client

select …
select …
Community Training
23
Lehrstuhl Informatik III: Datenbanksysteme
Prototype Implementation
Java-based prototype
 FreePastry library (Pastry implementation)




Presentations at



Rice University
MPI-SWS
BTW 2007, Aachen, Germany
VLDB 2007, Vienna, Austria
Deployed in various settings


LAN
WAN (AstroGrid-D, PlanetLab)
Community Training
24
Lehrstuhl Informatik III: Datenbanksysteme
Summary
Training phase
 Community-driven Data Grids



Domain-specific partitioning scheme
Partitioning scheme supports


Data skew
Region-based queries
Framework for comparing partitioning
schemes
 Various measures with regard to data load
balancing

Community Training
25
Lehrstuhl Informatik III: Datenbanksysteme
Ongoing Work

Database-driven comparison


0.1% of a Petabyte still is 1 Terabyte
Feasibility of median-based techniques
Workload-aware data partitioning
 Heterogeneous data nodes

Community Training
26
Lehrstuhl Informatik III: Datenbanksysteme
Get in Touch

Database systems group, TU München



HiSbase


Web site: http://www-db.in.tum.de
E-mail: [email protected]
http://www-db.in.tum.de/research/projects/hisbase/
Data stream management


“Grid-based Data Stream Processing in e-Science”
(e-Science '06)
http://www.gac-grid.de/project-products/Software/
DataStreamManagement.html
Community Training
27
Lehrstuhl Informatik III: Datenbanksysteme
AstroGrid-D Research Demo
Finding Galaxy Clusters using Grid Computing
Technology
 Room:
Banquet
 Wednesday
3:30 pm to
6:00 pm

Community Training
28