Lehrstuhl Informatik III: Datenbanksysteme Community Training: Partitioning Schemes in Good Shape...

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 ComputingBangalore, IndiaDecember 10th – 13th 2007

2Community Training

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

3Community Training

Lehrstuhl Informatik III: Datenbanksysteme

The Multiwavelength Milky Way

http://adc.gsfc.nasa.gov/mw/

4Community Training

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

5Community Training

Lehrstuhl Informatik III: Datenbanksysteme

Up-coming Data-intensive Applications

LOFAR

LHC

Data rates Terabytes a day/night Petabytes a year

LOFAR LSST Pan-STARRS LHC

6Community Training

Lehrstuhl Informatik III: Datenbanksysteme

Current Sharing in Data Grids

Images from: “Data-Intensive Grid Computing” by Srikumar Venugopal

Data autonomy Policies allow partners to access data Each institution ensures

Availability (replication) Scalability

Various organizational structures: Centralized Hierarchical Federated Hybrid

7Community Training

Lehrstuhl Informatik III: Datenbanksysteme

Community-Driven Data Distribution

8Community Training

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

9Community Training

Lehrstuhl Informatik III: Datenbanksysteme

Quadtrees

Well-known index structure

Recursive decomposition

Adaptive to data resolution

10Community Training

Lehrstuhl Informatik III: Datenbanksysteme

Quadtree-based Schemes: Splitting Variants

Center splitting Always bisects each

dimension congruent subareas Splitting points stored

or computed

Median heuristics Compute median for

each dimension independently

O(n) median algorithm

Splitting points stored

11Community Training

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

points-in-region queries self-match, cross-match queries

Equi-distant partitioning Declination coordinate Zone(ra, dec) = floor((dec + 90.0) / h)

Implemented directly in SQLImage by J. Gray et al.

12Community Training

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)

13Community Training

Lehrstuhl Informatik III: Datenbanksysteme

Skewed Training Data (Dskew

)

14Community Training

Lehrstuhl Informatik III: Datenbanksysteme

Comparing Partitioning Schemes

Duration Average data population Variance in partition population Empty partitions Size of the training set Baseline comparison

15Community Training

Lehrstuhl Informatik III: Datenbanksysteme

Average Data Population

0

10

20

30

40

50

60

70

80

90

100

1310726553632768163848192409610242566416

aver

age

data

pop

ulat

ion

(%)

# partitions

center splitting (skewed data)median splitting (skewed data)center splitting (uniform data)

median splitting (uniform data)

10% training sample, Dskew

avg(# objects in partitions)

# objects in biggest

partition

16Community Training

Lehrstuhl Informatik III: Datenbanksysteme

Evolution of the Partitioning Scheme

4096 partitions 8192 partitions 16384 partitions

17Community Training

Lehrstuhl Informatik III: Datenbanksysteme

Empty Leaves

0

1

2

3

4

5

6

7

8

1310726553632768163848192409610242566416

em

pty

pa

rtiti

on

s (%

)

# partitions

center splittingmedian splittingzones

10% training sample, Dskew

18Community Training

Lehrstuhl Informatik III: Datenbanksysteme

Size of the Training Set

1

4

16

64

256

1024

4096

16384

1310726553632768163848192409610242566416 0

5

10

15

20

25

30

35

sam

ple

ra

tio

em

pty

pa

rtiti

ons

(%)

# partitions

sample ratiocenter splitting

median splitting

0.1% training sample, Dskew

# objects in training set

# partitions to be created

19Community Training

Lehrstuhl Informatik III: Datenbanksysteme

Standard Quadtree vs. Median Heuristics

20Community Training

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

21Community Training

Lehrstuhl Informatik III: Datenbanksysteme

HiSbase

Peer-to-Peer layer assigns data partitions to peers

Higher flexibility New peers are

integrated seamlessly

22Community Training

Lehrstuhl Informatik III: Datenbanksysteme

Query Submission (at Peer d)

Determine relevant regions: [1,3]

Select coordinator:Region 1

Send CoordinateQuery-message to id1:{[1,3], SQL}

Message gets routed to Peer a.

01 2

3 4

5 6

23Community Training

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 …

24Community Training

Lehrstuhl Informatik III: Datenbanksysteme

Prototype Implementation

Java-based prototype FreePastry library (Pastry implementation)

Rice University MPI-SWS

Presentations at BTW 2007, Aachen, Germany VLDB 2007, Vienna, Austria

Deployed in various settings LAN WAN (AstroGrid-D, PlanetLab)

25Community Training

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

26Community Training

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

27Community Training

Lehrstuhl Informatik III: Datenbanksysteme

Get in Touch

Database systems group, TU München Web site: http://www-db.in.tum.de E-mail: scholl@in.tum.de

HiSbase 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

28Community Training

Lehrstuhl Informatik III: Datenbanksysteme

AstroGrid-D Research Demo

Finding Galaxy Clusters using Grid Computing Technology

Room:Banquet

Wednesday3:30 pm to6:00 pm