Graph lecture

An Overview of Graph Data Management and Analysis

M. Tamer Ozsu

University of WaterlooDavid R. Cheriton School of Computer Science

© M. Tamer Ozsu Croucher ASI (2015/12/16-18) 1 / 96

Graph Data are Very Common

Internet



Socialnetworks



Trade volumesand

connections



Biologicalnetworks



As of September 2011

MusicBrainz

(zitgist)

P20

Turismo de

Zaragoza

yovisto

Yahoo! Geo

Planet

YAGO

World Fact-book

El ViajeroTourism

WordNet (W3C)

WordNet (VUA)

VIVO UF

VIVO Indiana

VIVO Cornell

VIAF

URIBurner

Sussex Reading

Lists

Plymouth Reading

Lists

UniRef

UniProt

UMBEL

UK Post-codes

legislationdata.gov.uk

Uberblic

UB Mann-heim

TWC LOGD

Twarql

transportdata.gov.

uk

Traffic Scotland

theses.fr

Thesau-rus W

totl.net

Tele-graphis

TCMGeneDIT

TaxonConcept

Open Library (Talis)

tags2con delicious

t4gminfo

Swedish Open

Cultural Heritage

Surge Radio

Sudoc

STW

RAMEAU SH

statisticsdata.gov.

uk

St. Andrews Resource

Lists

ECS South-ampton EPrints

SSW Thesaur

us

SmartLink

Slideshare2RDF

semanticweb.org

SemanticTweet

Semantic XBRL

SWDog Food

Source Code Ecosystem Linked Data

US SEC (rdfabout)

Sears

Scotland Geo-

graphy

ScotlandPupils &Exams

Scholaro-meter

WordNet (RKB

Explorer)

Wiki

UN/LOCODE

Ulm

ECS (RKB

Explorer)

Roma

RISKS

RESEX

RAE2001

Pisa

OS

OAI

NSF

New-castle

LAASKISTI

JISC

IRIT

IEEE

IBM

Eurécom

ERA

ePrints dotAC

DEPLOY

DBLP (RKB

Explorer)

Crime Reports

UK

Course-ware

CORDIS (RKB

Explorer)CiteSeer

Budapest

ACM

riese

Revyu

researchdata.gov.

ukRen. Energy Genera-

tors

referencedata.gov.

uk

Recht-spraak.

nl

RDFohloh

Last.FM (rdfize)

RDF Book

Mashup

Rådata nå!

PSH

Product Types

Ontology

ProductDB

PBAC

Poké-pédia

patentsdata.go

v.uk

OxPoints

Ord-nance Survey

Openly Local

Open Library

OpenCyc

Open Corpo-rates

OpenCalais

OpenEI

Open Election

Data Project

OpenData

Thesau-rus

Ontos News Portal

OGOLOD

JanusAMP

Ocean Drilling Codices

New York

Times

NVD

ntnusc

NTU Resource

Lists

Norwe-gian

MeSH

NDL subjects

ndlna

myExperi-ment

Italian Museums

medu-cator

MARC Codes List

Man-chester Reading

Lists

Lotico

Weather Stations

London Gazette

LOIUS

Linked Open Colors

lobidResources

lobidOrgani-sations

LEM

LinkedMDB

LinkedLCCN

LinkedGeoData

LinkedCT

LinkedUser

FeedbackLOV

Linked Open

Numbers

LODE

Eurostat (OntologyCentral)

Linked EDGAR

(OntologyCentral)

Linked Crunch-

base

lingvoj

Lichfield Spen-ding

LIBRIS

Lexvo

LCSH

DBLP (L3S)

Linked Sensor Data (Kno.e.sis)

Klapp-stuhl-club

Good-win

Family

National Radio-activity

JP

Jamendo (DBtune)

Italian public

schools

ISTAT Immi-gration

iServe

IdRef Sudoc

NSZL Catalog

Hellenic PD

Hellenic FBD

PiedmontAccomo-dations

GovTrack

GovWILD

GoogleArt

wrapper

gnoss

GESIS

GeoWordNet

GeoSpecies

GeoNames

GeoLinkedData

GEMET

GTAA

STITCH

SIDER

Project Guten-berg

MediCare

Euro-stat

(FUB)

EURES

DrugBank

Disea-some

DBLP (FU

Berlin)

DailyMed

CORDIS(FUB)

Freebase

flickr wrappr

Fishes of Texas

Finnish Munici-palities

ChEMBL

FanHubz

EventMedia

EUTC Produc-

tions

Eurostat

Europeana

EUNIS

EU Insti-

tutions

ESD stan-dards

EARTh

Enipedia

Popula-tion (En-AKTing)

NHS(En-

AKTing) Mortality(En-

AKTing)

Energy (En-

AKTing)

Crime(En-

AKTing)

CO2 Emission

(En-AKTing)

EEA

SISVU

education.data.g

ov.uk

ECS South-ampton

ECCO-TCP

GND

Didactalia

DDC Deutsche Bio-

graphie

datadcs

MusicBrainz

(DBTune)

Magna-tune

John Peel

(DBTune)

Classical (DB

Tune)

AudioScrobbler (DBTune)

Last.FM artists

(DBTune)

DBTropes

Portu-guese

DBpedia

dbpedia lite

Greek DBpedia

DBpedia

data-open-ac-uk

SMCJournals

Pokedex

Airports

NASA (Data Incu-bator)

MusicBrainz(Data

Incubator)

Moseley Folk

Metoffice Weather Forecasts

Discogs (Data

Incubator)

Climbing

data.gov.uk intervals

Data Gov.ie

databnf.fr

Cornetto

reegle

Chronic-ling

America

Chem2Bio2RDF

Calames

businessdata.gov.

uk

Bricklink

Brazilian Poli-

ticians

BNB

UniSTS

UniPathway

UniParc

Taxonomy

UniProt(Bio2RDF)

SGD

Reactome

PubMedPub

Chem

PRO-SITE

ProDom

Pfam

PDB

OMIMMGI

KEGG Reaction

KEGG Pathway

KEGG Glycan

KEGG Enzyme

KEGG Drug

KEGG Com-pound

InterPro

HomoloGene

HGNC

Gene Ontology

GeneID

Affy-metrix

bible ontology

BibBase

FTS

BBC Wildlife Finder

BBC Program

mes BBC Music

Alpine Ski

Austria

LOCAH

Amster-dam

Museum

AGROVOC

AEMET

US Census (rdfabout)

Media

Geographic

Publications

Government

Cross-domain

Life sciences

User-generated content

Linked data


Linking Open Data cloud diagram, by Richard Cyganiak and Anja Jentzsch.http://lod-cloud.net/

Outline

1 Introduction – Graph Types

2 Property Graph ProcessingClassificationOnline queryingOffline analytics

3 RDF Graph QueryingData WarehousingDistributed SPARQL ExecutionLinked Object Data Querying


Outline





Graph Types

Property graph

film 2014(initial release date, “1980-05-23”)

(label, “The Shining”)

books 0743424425(rating, 4.7)

offers 0743424425amazonOffer

geo 2635167(name, “United Kingdom”)

(population, 62348447) actor 29704(actor name, “Jack Nicholson”)

film 3418(label, “The Passenger”)

film 1267(label, “The Last Tycoon”)

director 8476(director name, “Stanley Kubrick”)

film 2685(label, “A Clockwork Orange”)

film 424(label, “Spartacus”)

actor 30013

(relatedBook)

(hasOffer)

(based near)(actor)

(director) (actor)

(actor) (actor)

(director) (director)


Graph Types

RDF graph

mdb:film/2014

“1980-05-23”

movie:initial release date

“The Shining”refs:label

bm:books/0743424425

4.7

rev:rating

bm:offers/0743424425amazonOffer

geo:2635167

“United Kingdom”

gn:name

62348447

gn:population

mdb:actor/29704

“Jack Nicholson”

movie:actor name

mdb:film/3418

“The Passenger”

refs:label

mdb:film/1267

“The Last Tycoon”

refs:label

mdb:director/8476

“Stanley Kubrick”

movie:director name

mdb:film/2685

“A Clockwork Orange”

refs:label

mdb:film/424

“Spartacus”

refs:label

mdb:actor/30013

movie:relatedBook

scam:hasOffer

foaf:based nearmovie:actor

movie:directormovie:actor

movie:actor movie:actor

movie:director movie:director


Graph Types

Property graph



books 0743424425(rating, 4.7)









actor 30013

(relatedBook)

(hasOffer)

(based near)(actor)

(director) (actor)

(actor) (actor)


Workload: Online queries andanalytic workloads

Query execution: Varies

RDF graph

mdb:film/2014

“1980-05-23”



bm:books/0743424425

4.7

rev:rating


geo:2635167


gn:name

62348447

gn:population

mdb:actor/29704


movie:actor name

mdb:film/3418

“The Passenger”

refs:label

mdb:film/1267


refs:label

mdb:director/8476


movie:director name

mdb:film/2685


refs:label

mdb:film/424

“Spartacus”

refs:label

mdb:actor/30013

movie:relatedBook

scam:hasOffer





Workload: SPARQL queries

Query execution: subgraphmatching by homomorphism


Outline





Outline





Classification [Ammar and Ozsu, 2015]

Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online

Streaming Incremental

Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Focus here is on the

dynamism of the

graphs in whether or

not they change and

how they change.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads


dynamism of the


not they change and

how they change.


how algorithms behave

as their input changes.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads


dynamism of the


not they change and

how they change.


how algorithms behave

as their input changes.

The types of workloads

that the approaches are

designed to handle.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Graphs do not

change or we

are not inter-

ested in their

changes – only

a snapshot is

considered.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Graphs do not

change or we

are not inter-

ested in their

changes – only

a snapshot is

considered.

Graphs change

and we are

interested in

their changes.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Graphs do not

change or we

are not inter-

ested in their

changes – only

a snapshot is

considered.

Graphs change

and we are

interested in

their changes.

Dynamic

graphs with

high veloc-

ity changes –

not possible to

see the entire

graph at once.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Graphs do not

change or we

are not inter-

ested in their

changes – only

a snapshot is

considered.

Graphs change

and we are

interested in

their changes.

Dynamic

graphs with

high veloc-

ity changes –

not possible to

see the entire

graph at once.

Dynamic

graphs with un-

known changes

– requires re-

discovery of

the graph (e.g.,

LOD).



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Computation accesses a

portion of the graph

and the results are

computed for a subset

of vertices; e.g., point-

to-point shortest path,

subgraph matching,

reachability, SPARQL.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Computation accesses a

portion of the graph

and the results are

computed for a subset

of vertices; e.g., point-

to-point shortest path,

subgraph matching,

reachability, SPARQL.

Computation accesses

the entire graph and

may require multiple

iterations; e.g., PageR-

ank, clustering, graph

colouring, all pairs

shortest path.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Sees the en-

tire input in

advance.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Sees the en-

tire input in

advance.

Sees the input

piece-meal as it

executes.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Sees the en-

tire input in

advance.

Sees the input

piece-meal as it

executes.

One-pass on-

line algorithm

with limited

memory.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Sees the en-

tire input in

advance.

Sees the input

piece-meal as it

executes.

One-pass on-

line algorithm

with limited

memory.

Online algo-

rithm with

some info

about forth-

coming input.© M. Tamer Ozsu Croucher ASI (2015/12/16-18) 8 / 96


Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Sees the en-

tire input in

advance.

Sees the input

piece-meal as it

executes.

One-pass on-

line algorithm

with limited

memory.

Online algo-

rithm with

some info

about forth-

coming input.

Sees the en-

tire input

in advance,

which may

change; an-

swers computed

as change oc-

curs.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Sees the en-

tire input in

advance.

Sees the input

piece-meal as it

executes.

One-pass on-

line algorithm

with limited

memory.

Online algo-

rithm with

some info

about forth-

coming input.

Sees the en-

tire input

in advance,

which may

change; an-

swers computed

as change oc-

curs.

Similar to dy-

namic, but

computation

happens in

batches of

changes.© M. Tamer Ozsu Croucher ASI (2015/12/16-18) 8 / 96

Example Design Points

Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Compute the query result/perform analytic computation over the graphas it exists.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Compute the query result/perform analytic computation over the graphas it is revealed.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Compute the query result/perform analytic computation on each snap-shot from scratch.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Continuously compute the query result/perform analytic computation asthe input changes.



Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Compute the query result/perform analytic computation after a batch ofinput changes.


Example Design Points – Not all alternatives make sense

Graph Dynamism

StaticGraphs

DynamicGraphs

StreamingGraphs

EvolvingGraphs

Algorithm Types

Offline Online


Dynamic

BatchDynamic

Workload Types

OnlineQueries

AnalyticsWorkloads

Dynamic (or batch-dynamic) algorithms do not make sense for staticgraphs.


Graph Processing Systems

System Memory/Disk

ArchitectureComputingparadigm

SupportedWorkloads

Hadoop Disk Parallel/Distributed MapReduce Analytical

Haloop Disk Parallel/Distributed MapReduce Analytical

Pegasus Disk Parallel/Distributed MapReduce Analytical

GraphX Disk Parallel/DistributedMapReduce

(Spark)Analytical

Pregel/Giraph Memory Parallel/Distributed Vertex-Centric Analytical

GraphLab Memory Parallel/Distributed Vertex-Centric Analytical

GraphChi Disk Single machine Vertex-Centric Analytical

Stream Disk Single machine Edge-Centric Analytical

Trinity Memory Parallel/DistributedFlexible using K-V

store on DSMOnline &Analytical

Titan Disk Parallel/DistributedK-V store

(Cassandra)Online

Neo4J Disk Single machineProcedural/Linked-list

Online


Graph Workloads

Online graph querying

Reachability

Single source shortest-path

Subgraph matching

SPARQL queries

Offline graph analytics

PageRank

Clustering

Strongly connectedcomponents

Diameter finding

Graph colouring

All pairs shortest path

Graph pattern mining

Machine learning algorithms(Belief propagation, Gaussiannon-negative matrixfactorization)


Outline





Reachability Queries



books 0743424425(rating, 4.7)









actor 30013

(relatedBook)

(hasOffer)

(based near)(actor)

(director) (actor)

(actor) (actor)






books 0743424425(rating, 4.7)









actor 30013

(relatedBook)

(hasOffer)

(based near)(actor)

(director) (actor)

(actor) (actor)


Can you reach film 1267 from film 2014?





books 0743424425(rating, 4.7)









actor 30013

(relatedBook)

(hasOffer)

(based near)(actor)

(director) (actor)

(actor) (actor)


Is there a book whose rating is > 4.0 associated with a film that wasdirected by Stanley Kubrick?



Think of Facebook graph and finding friends of friends.


Subgraph Matching

?m ?dmovie:director

?name

rdfs:label

?b

movie:relatedBook


movie:director name

?rrev:rating

FILTER(?r > 4.0)

mdb:film/2014

“1980-05-23”



bm:books/0743424425

4.7

rev:rating


geo:2635167


gn:name

62348447

gn:population

mdb:actor/29704


movie:actor name

mdb:film/3418

“The Passenger”

refs:label

mdb:film/1267


refs:label

mdb:director/8476


movie:director name

mdb:film/2685


refs:label

mdb:film/424

“Spartacus”

refs:label

mdb:actor/30013

movie:relatedBook

scam:hasOffer





SubgraphM

atching


Outline





PageRank Computation

A web page is important if it is pointed to by other importantpages.

P1 P2

P3

P5P6

P4

r(Pi ) =∑

Pj∈BPi

r(Pj)

|FPj|

r(P2) =r(P1)

2+

r(P3)

3

rk+1(Pi ) =∑

Pj∈BPi

rk(Pj)

|FPj|

BPi: in-neighbours of Pi

FPi: out-neighbours of Pi


PageRank Computation

A web page is important if it is pointed to by other importantpages.

P1 P2

P3

P5P6

P4

rk+1(Pi ) =∑

Pj∈BPi

rk(Pj)

|FPj|

Iteration 0 Iteration 1 Iteration 2Rank atIter. 2

r0(P1) = 1/6 r1(P1) = 1/18 r2(P1) = 1/36 5r0(P2) = 1/6 r1(P2) = 5/36 r2(P2) = 1/18 4r0(P3) = 1/6 r1(P3) = 1/12 r2(P3) = 1/36 5r0(P4) = 1/6 r1(P4) = 1/4 r2(P4) = 17/72 1r0(P5) = 1/6 r1(P5) = 5/36 r2(P5) = 11/72 3r0(P6) = 1/6 r1(P6) = 1/6 r2(P6) = 14/72 2

Iterative processing.


Some Alternative Computational Models for OfflineAnalytics

Vertex-centric (Scatter-Gather)Specify (a) computation at each vertex, and (b) communication withneighbour verticesSynchronous – Pregel [Malewicz et al., 2010], GiraphAsynchronous – GraphLab [Low et al., 2012]

Block-centricSimilar to vertex-centric but on blocks for communication

Connected subgraph of the graph

Blogel [Yan et al., 2014]MapReduce

Need to save in HDFS intermediate results of each iteration – bothgood and badHadoop, Haloop [Bu et al., 2012]

Modified MapReduceBased on Spark [Zaharia et al., 2010; Zaharia, 2016]

Keep intermediate states in memoryProvide fault-tolerance by keeping lineage

GraphX [Gonzalez et al., 2014]






Blogel [Yan et al., 2014]

MapReduceNeed to save in HDFS intermediate results of each iteration – bothgood and badHadoop, Haloop [Bu et al., 2012]

























Vertex-Centric Computation

“Think like a vertex”

vertex_scatter(vertex v)

Push local computation toneighbours on the out-boundedges

vertex_gather(vertex v)

Gather local computation fromneighbours on the in-bound edges

Continue until all vertices areinactive

Vertex state machine

?

Active Inactive

Vote halt

Message received


Vertex-Centric Computation

“Think like a vertex”

vertex_scatter(vertex v)

Push local computation toneighbours on the out-boundedges

vertex_gather(vertex v)

Gather local computation fromneighbours on the in-bound edges

Continue until all vertices areinactive

Vertex state machine

?

Active Inactive

Vote halt

Message received


Synchronous Vertex-Centric Computation

Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3

CommunicationBarrier

Each machine performsvertex-centric computationon its graph partition


Superstep 1 Superstep 2 Superstep 3

Computation



Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3





Computation



Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3





Computation



Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3





Computation



Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3





Computation


Asynchronous Vertex-Centric Computation

No communication barriers. 3

Uses the most recent vertex values. 3

Implemented via distributed locking

Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3

v0

v1 v2

v3 v4






Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3

v0

v1 v2

v3 v4






Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3

v0

v1 v2

v3 v4






Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3

v0

v1 v2

v3 v4






Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3

v0

v1 v2

v3 v4






Machine 1

Machine 2

Machine 3

Machine 1

Machine 2

Machine 3

v0

v1 v2

v3 v4


Summary of an Experiment [Han et al., 2014]

A large study comparing Giraph, GraphLab, GPS, Mizan.

1 Giraph scales better across graphs;GraphLab scales better across more machines.

2 Distributed locking for asynchronous execution is not scalable –Performance degrades as more machines are used due to lockcontention, termination scheme, lack of message batching

3 Graph storage should be memory and mutation efficient.

4 Message processing optimizations are very important.

5 Workloads have different resource demands









64 machines TW UK

Giraph (byte array) 5.8GB 7.0GBGraphLab (sync) 4.5GB 14GB

TW 16 machines 128 machines

Giraph (byte array) 8.5GB 5.8GBGraphLab (sync) 11GB 3.3GB

























No Mutations

Time Memory

Byte array 3 3Hash map 7 7

With Mutations (DMST)

Time Memory

Byte array 77 3Hash map 3 7











A large study comparing Giraph, GraphLab, GPS, Mizan.1 Giraph scales better across graphs;

GraphLab scales better across more machines.2 Distributed locking for asynchronous execution is not scalable –

Performance degrades as more machines are used due to lockcontention, termination scheme, lack of message batching

3 Graph storage should be memory and mutation efficient.4 Message processing optimizations are very important.5 Workloads have different resource demands

Algorithm CPU Memory Network

PageRank Medium Medium HighSSSP Low Low LowWCC Low Medium MediumDMST High High Medium


Block-Centric Computation

Blogel [Yan et al., 2014]: “Think like a block”; also “think like agraph” [Tian et al., 2013]

Vertex-centric assumes all vertices communicate over the network;this is not efficient

Read-world graphs have skewed vertex degree distribution

Common in power-law graphsProblem: imbalanced communication workloads

Real-world graphs have large diameters

Common in road networks, web graphs, terrain meshesProblem: one superstep per hop ⇒ too many supersteps

Real-world graphs have high average vertex degree

Common in social networks, mobile communication networksProblem: heavy average communication workloads


Blogel Principles

Exploit the partitioning of the graph

Message exchanges only among blocks

Block: a connected subgraph of the graph

Within a block, run a serial in-memory algorithm; no need to follow aBSP model


Benefits of Block-Centric Computation

High-degree vertices inside a block send no messages

Fewer number of supersteps

Fewer number of blocks than vertices


Example: Weakly Connected Component

Algorithm exchanges vertex id’swith neighbours

id(vi )← min{vi , vj , . . . , vk}where vj , . . . , vk are neighboursof vi

Vertex-centric requires everyvertex sends to its neighboursuntil every vertex is reached

Block-centric needs twoiterations:

1 All vertices in partition Aexchange ids; X and Y sendids to neighbours in partitionB

2 All vertices in partition Bexchange ids

A B

0

X

Y


Block Construction

The partitioning algorithm needs to maximize number of vertices thathave all their edges in the same partition

Hash partitioning is not suitable because many vertices will probablyhave at least one cut-edge

URL partitioner

For web graphs: based on domain names of web page nodes

2D partitioner

For spatial networks: based on coordinates of node

Graph Voronoi diagram partitioner

For general graphs


MapReduce Basics [Li et al., 2014]

For data analysis of very large data sets

Highly dynamic, irregular, schemaless, etc.SQL too heavy

“Embarrassingly parallel problems”

New, simple parallel programming modelData structured as (key, value) pairs

E.g. (doc-id, content), (word, count), etc.

Functional programming style with two functions to be given:

Map(k1,v1) → list(k2,v2)

Reduce(k2, list (v2)) → list(v3)

Implemented on a distributed file system (e.g., Google File System)on very large clusters


MapReduce Processing

...Inp

ut

dat

ase

t

Map

Map

Map

Map

(k1, v)

(k2, v)(k2, v)

(k2, v)

(k1, v)

(k1, v)

(k2, v)

Group by k

Group by k

(k1, (v , v , v))

(k1, (v , v , v , v)) Reduce

Reduce

Ou

tpu

td

ata

set


MapReduce Architecture

Scheduler

Master

Input Module

Map Module

Combine Module

Partition Module

Map Process

Worker

Input Module

Map Module

Combine Module

Partition Module

Map Process

Worker

Input Module

Map Module

Combine Module

Partition Module

Map Process

Worker

Group Module

Reduce Module

Output Module

Reduce Process

Worker

Group Module

Reduce Module

Output Module

Reduce Process

Worker


Execution Flow with Architecture [Dean and Ghemawat, 2008]MapReduce: Simplified Data Processing on Large Clusters

7. When all map tasks and reduce tasks have been completed, the mas-ter wakes up the user program. At this point, the MapReduce callin the user program returns back to the user code.

After successful completion, the output of the mapreduce executionis available in the R output files (one per reduce task, with file namesspecified by the user). Typically, users do not need to combine these Routput files into one file; they often pass these files as input to anotherMapReduce call or use them from another distributed application thatis able to deal with input that is partitioned into multiple files.

3.2 Master Data StructuresThe master keeps several data structures. For each map task andreduce task, it stores the state (idle, in-progress, or completed) and theidentity of the worker machine (for nonidle tasks).

The master is the conduit through which the location of interme-diate file regions is propagated from map tasks to reduce tasks. There -fore, for each completed map task, the master stores the locations andsizes of the R intermediate file regions produced by the map task.Updates to this location and size information are received as map tasksare completed. The information is pushed incrementally to workersthat have in-progress reduce tasks.

3.3 Fault ToleranceSince the MapReduce library is designed to help process very largeamounts of data using hundreds or thousands of machines, the librarymust tolerate machine failures gracefully.

Handling Worker FailuresThe master pings every worker periodically. If no response is receivedfrom a worker in a certain amount of time, the master marks the workeras failed. Any map tasks completed by the worker are reset back to theirinitial idle state and therefore become eligible for scheduling on otherworkers. Similarly, any map task or reduce task in progress on a failedworker is also reset to idle and becomes eligible for rescheduling.

Completed map tasks are reexecuted on a failure because their out-put is stored on the local disk(s) of the failed machine and is thereforeinaccessible. Completed reduce tasks do not need to be reexecutedsince their output is stored in a global file system.

When a map task is executed first by worker A and then later exe-cuted by worker B (because A failed), all workers executing reducetasks are notified of the reexecution. Any reduce task that has notalready read the data from worker A will read the data from worker B.

MapReduce is resilient to large-scale worker failures. For example,during one MapReduce operation, network maintenance on a runningcluster was causing groups of 80 machines at a time to become unreach-able for several minutes. The MapReduce master simply re executed thework done by the unreachable worker machines and continued to makeforward progress, eventually completing the MapReduce operation.

Semantics in the Presence of FailuresWhen the user-supplied map and reduce operators are deterministicfunctions of their input values, our distributed implementation pro-duces the same output as would have been produced by a nonfaultingsequential execution of the entire program.

split 0

split 1

split 2

split 3

split 4

(1) fork

(3) read(4) local write

(1) fork(1) fork

(6) write

worker

worker

worker

Master

UserProgram

outputfile 0

outputfile 1

worker

worker

(2)assignmap

(2)assignreduce

(5) remote

(5) read

Inputfiles

Mapphasr

Intermediate files(on local disks)

Reducephase

Outputfiles

Fig. 1. Execution overview.

COMMUNICATIONS OF THE ACM January 2008/Vol. 51, No. 1 109


Hadoop

Most popular MapReduce implementation – developed by Yahoo!Two components

Processing engineHDFS: Hadoop Distributed Storage System – others possibleCan be deployed on the same machine or on different machines

ProcessesJob tracker: hosted on the master node and implements the scheduleTask tracker: hosted on the worker nodes and accepts tasks from job trackerand executes them

HDFSName node: stores how data are partitioned, monitors the status of datanodes, and data dictionaryData node: Stores and manages data chunks assigned to it

Task Tracker Job Tracker Task Tracker

Data Node Name Node Data Node

Worker 1 Name Node Worker n

MapReduce

HDFS


HaLoop [Bu et al., 2012]

Overcome MapReduce shortcomings for iterative jobs

Having to save data in HDFS in between each iterationChecking the fixpoint requires a new job at each iteration

Scheduler change: assign to the same machine the map & reducetasks that occur in different iterations but access the same data

Cache invariant data

Cache reduce output to easily check for fixpoint


Spark System

MapReduce does not perform well in iterative computations

Workflow model is acyclicHave to write to HDFS after each iteration and have to read fromHDFS at the beginning of next iteration

Spark objectives

Better support for iterative programsProvide a complete ecosystemSimilar abstraction (to MapReduce) for programmingMaintain MapReduce fault-tolerance and scalability

Fundamental concepts

RDD: Reliable Distributed DatasetsCaching of working setMaintaining lineage for fault-tolerance


Spark System



Spark objectives





Spark System



Spark objectives





Spark Ecosystem [Michiardi, 2015]

NativeSparkApps

SparkSQL

SparkStreaming

MLlib(machinelearning)

GraphX(graph

processing)

Apache Spark


Spark Programming Model [Zaharia et al., 2010, 2012]

HDFS

Create RDD

· · ·

RDD

Cache? CacheYes

TransformRDD?

No

Process

No

TransformYes

HDFS

Each transform generates anew RDD that may also becached or processed

Created from HDFS or parallelized arrays;Partitioned across worker machines;May be made persistent lazily;

Processing done on one of the RDDs;Done in parallel across workers;First processing on a RDD is from disk;Subsequent processing of the same RDD from cache



HDFS

Create RDD

· · ·

RDD

Cache? CacheYes

TransformRDD?

No

Process

No

TransformYes

HDFS






HDFS

Create RDD

· · ·

RDD

Cache? CacheYes

TransformRDD?

No

Process

No

TransformYes

HDFS






HDFS

Create RDD

· · ·

RDD

Cache? CacheYes

TransformRDD?

No

Process

No

TransformYes

HDFS





Example – Log Mining [Zaharia et al., 2010, 2012]

Load log messages from a file system, create a new file by filtering theerror messages, read this file into memory, then interactively search forvarious patterns

lines = spark.textFile(hdfs://...)

CreateRDD

errors = lines.filter( .startsWith(ERROR))

Transform RDD

messages = errors.map( .split(‘\t ’)(2))

Another transform

cachedMsgs = messages.cache()

Cache results

cachedMsgs.filter( .contains(foo)).count

Action

cachedMsgs.filter( .contains(bar)).count

Another Action

accesses cache

Driver

WorkerWorkerWorker

Block 1 Block 2 Block 3

TasksResults

Cache Cache Cache



Load log messages from a file system, create a new file by filtering theerror messages, read this file into memory, then interactively search forvarious patternslines = spark.textFile(hdfs://...)

CreateRDD


Transform RDD


Another transform


Cache results


Action


Another Action

accesses cache

Driver

WorkerWorkerWorker


TasksResults

Cache Cache Cache




CreateRDD


Transform RDD


Another transform


Cache results


Action


Another Action

accesses cache

Driver

WorkerWorkerWorker


TasksResults

Cache Cache Cache




CreateRDD


Transform RDD


Another transform


Cache results


Action


Another Action

accesses cache

Driver

WorkerWorkerWorker


TasksResults

Cache Cache Cache




CreateRDD


Transform RDD


Another transform


Cache results


Action


Another Action

accesses cache

Driver

WorkerWorkerWorker


TasksResults

Cache Cache Cache




CreateRDD


Transform RDD


Another transform


Cache results


Action


Another Action

accesses cache

Driver

WorkerWorkerWorker


TasksResults

Cache Cache Cache




CreateRDD


Transform RDD


Another transform


Cache results


Action


Another Action

accesses cache

Driver

WorkerWorkerWorker


Tasks

Results

Cache Cache Cache




CreateRDD


Transform RDD


Another transform


Cache results


Action


Another Action

accesses cache

Driver

WorkerWorkerWorker


TasksResults

Cache Cache Cache




CreateRDD


Transform RDD


Another transform


Cache results


Action


Another Action

accesses cache

Driver

WorkerWorkerWorker


TasksResults

Cache Cache Cache




CreateRDD


Transform RDD


Another transform


Cache results


Action


Another Action

accesses cache

Driver

WorkerWorkerWorker


TasksResults

Cache Cache Cache


RDD and Processing

HDFS

lines = spark.textFile(hdfs://...)

linesError, msg1

Warn, msg2

Error, msg1

Info, msg8

Warn, msg2

Info, msg8

Error, msg3

Info, msg5

Info, msg5

Error, msg4

Warn, msg9

Error, msg1

errors


Error, msg1

Error, msg1

Error, msg3 Error, msg4

Error, msg1

messages

messages = errors.map .split(‘\t ’)(2)

msg1

msg1

msg3 msg4

msg1

Th

ese

are

no

tye

tg

ener

ated


RDD and Processing

lineserrors

messagesmsg1

msg1

msg3 msg4

msg1

lines

messages.filter( .contains(foo)).count

errors

messagesmsg1

msg1

msg3 msg4

msg1

Now

the

RD

Ds

are

mat

eria

lized

;

Co

mm

and

no

tye

tex

ecu

ted

Driver

messages.filter( .contains(foo)).count



Built on top of Spark

Objective is to combine data analytics with graph processing

Unify computation on tables and graphs

Carefully convert graph to tabular representation

Native GraphX API or can accommodate vertex-centric computation

NativeSparkApps

SparkSQL

SparkStreaming

MLlib(machinelearning)

GraphX(graph

processing)

Apache Spark

Vertex-centric API

AppApp

App App


GraphX: Representation of Graphs as Tables

A

B

C

D

E

F

G

H

I

J



Partition 1

Partition 2

A

B

C

D

E

F

G

H

I

J

Edge-disjointpartitioning



Partition 1

Partition 2

Mac

hin

e1

Mac

hin

e2

Vertex Table

(RDD)v-prop:vertex prop.

A

B

C

D

E

F

G

H

I

J


A v-prop

B v-prop

...

I v-prop

D v-prop

E v-prop

F v-prop

J v-prop



Partition 1

Partition 2

Mac

hin

e1

Mac

hin

e2

Vertex Table


Edge Table

(RDD)e-prop:edge prop.

A

B

C

D

E

F

G

H

I

J


A v-prop

B v-prop

...

I v-prop

D v-prop

E v-prop

F v-prop

J v-prop

A e-prop B

A e-prop C

...

F e-prop G

A e-prop D

A e-prop E...

E e-prop F



Partition 1

Partition 2

Mac

hin

e1

Mac

hin

e2

Vertex Table


Edge Table


A

B

C

D

E

F

G

H

I

J


A v-prop

B v-prop

...

I v-prop

D v-prop

E v-prop

F v-prop

J v-prop

A e-prop B

A e-prop C

...

F e-prop G

A e-prop D

A e-prop E...

E e-prop FJoining vertices

and edgesMove vertices to edges



Partition 1

Partition 2

Mac

hin

e1

Mac

hin

e2

Vertex Table


Edge Table


RoutingTable

(RDD)

A

B

C

D

E

F

G

H

I

J


A v-prop

B v-prop

...

I v-prop

D v-prop

E v-prop

F v-prop

J v-prop

A e-prop B

A e-prop C

...

F e-prop G

A e-prop D

A e-prop E...

E e-prop F

A 1 2

B 1

...

I 1

F 1 2

D 2

E 2

J 2


GraphX: Computation Model

Mac

hin

e1

Mac

hin

e2

Vertex Table Edge Table

A v-prop

B v-prop

...

I v-prop

D v-prop

E v-prop

F v-prop

J v-prop

A e-prop B

A e-prop C

...

F e-prop G

A e-prop D

A e-prop E...

E e-prop F



Mac

hin

e1

Mac

hin

e2


A v-prop

B v-prop

...

I v-prop

D v-prop

E v-prop

F v-prop

J v-prop

A e-prop B

A e-prop C

...

F e-prop G

A e-prop D

A e-prop E...

E e-prop F

First Phase: JoinVertex table on Edge table

Triples View

A v-prop e-prop B v-prop

A v-prop e-prop C v-prop

C v-prop e-prop G v-prop

...

E v-prop e-prop G v-prop

J v-prop e-prop G v-prop



Mac

hin

e1

Mac

hin

e2


A v-prop

B v-prop

...

I v-prop

D v-prop

E v-prop

F v-prop

J v-prop

A e-prop B

A e-prop C

...

F e-prop G

A e-prop D

A e-prop E...

E e-prop FTriples View

A v-prop e-prop B v-prop

A v-prop e-prop C v-prop

C v-prop e-prop G v-prop

...

E v-prop e-prop G v-prop

J v-prop e-prop G v-prop

Second Phase: Compute neighbourhoodGroup-by aggregate


GraphX: Operators

Table transform operators – inherited from Sparkmap(func) Return a new RDD formed by passing each element

of the source through a function func

filter(func) Return a new RDD formed by selecting thoseelements of the source on which func returns true

flatMap(func) Similar to map, but each input item can be mappedto 0 or more output items

mapPartitions(func) Similar to map, but runs separately on each partition(block) of the RDD, so func must be of type Iterator

sample(repl , fraction,seed)

Sample a fraction fraction of the data, with orwithout replacement (set repl accordingly), using agiven random number generator seed

union(otherDataset)intersection()

Return a new RDD containing the union/intersectionof the elements in the source RDD and the argument

groupByKey() Operates on a RDD of (K, V) pairs, returns a RDDof (K, Iterable<V>) pairs

reduceByKey(func, . . .) Operates on a RDD of (K, V) pairs, returns a RDDof (K, V) pairs where the values for each key areaggregated using the given reduce function func

Graph operatorsGraph(vertex coll ,edge coll)

Logically binds together a pair of vertex and edgeproperty collections into a property graph; verifiesthat each vertex occurs only once and edges connectexisting vertices

triplets(vertex coll ,vertex coll , edge coll)

Returns the triplets view of the graph

mrTriplets(map,reduce) MapReduce triplets - encodes the two-stage processof join to create triplets and group by


GraphX: Operators

Table transform operators – inherited from Spark

Graph operatorsGraph(vertex coll ,edge coll)

Logically binds together a pair of vertex and edgeproperty collections into a property graph; verifiesthat each vertex occurs only once and edges connectexisting vertices

triplets(vertex coll ,vertex coll , edge coll)

Returns the triplets view of the graph

mrTriplets(map,reduce) MapReduce triplets - encodes the two-stage processof join to create triplets and group by


Outline





RDF Introduction

Everything is an uniquely namedresource

Prefixes can be used to shorten thenames

Properties of resources can be defined

Relationships with other resources canbe defined

Resource descriptions can becontributed by different people/groupsand can be located anywhere in the web

Integrated web “database”

http://data.linkedmdb.org/resource/actor/JN29704

xmlns:y=http://data.linkedmdb.org/resource/actor/

y:JN29704

y:JN29704:hasName “Jack Nicholson”

y:JN29704:BornOnDate “1937-04-22”

y:TS2014:title “The Shining”

y:TS2014:releaseDate “1980-05-23”

y:TS2014

JN29704:movieActor


RDF Introduction









y:JN29704


y:JN29704:BornOnDate “1937-04-22”



y:TS2014

JN29704:movieActor


RDF Introduction









y:JN29704


y:JN29704:BornOnDate “1937-04-22”



y:TS2014

JN29704:movieActor


RDF Introduction









y:JN29704


y:JN29704:BornOnDate “1937-04-22”



y:TS2014

JN29704:movieActor


RDF Introduction









y:JN29704


y:JN29704:BornOnDate “1937-04-22”



y:TS2014

JN29704:movieActor


RDF Data Model

Triple: Subject, Predicate (Property), Object(s, p, o)

Subject: the entity that is described (URIor blank node)

Predicate: a feature of the entity (URI)Object: value of the feature (URI, blank

node or literal)

(s, p, o) ∈ (U ∪ B)× U × (U ∪ B ∪ L)

Set of RDF triples is called an RDF graph

U

Subject Object

U B U B L

U: set of URIsB: set of blank nodesL: set of literals

Predicate

Subject Predicate Objecthttp://...imdb.../film/2014 rdfs:label “The Shining”http://...imdb.../film/2014 movie:releaseDate “1980-05-23”http://...imdb.../29704 movie:actor name “Jack Nicholson”. . . . . . . . .


RDF Example InstancePrefixes: mdb=http://data.linkedmdb.org/resource/; geo=http://sws.geonames.org/

bm=http://wifo5-03.informatik.uni-mannheim.de/bookmashup/lexvo=http://lexvo.org/id/;wp=http://en.wikipedia.org/wiki/

Subject Predicate Object

mdb: film/2014 rdfs:label “The Shining”mdb:film/2014 movie:initial release date “1980-05-23”’mdb:film/2014 movie:director mdb:director/8476mdb:film/2014 movie:actor mdb:actor/29704mdb:film/2014 movie:actor mdb: actor/30013mdb:film/2014 movie:music contributor mdb: music contributor/4110mdb:film/2014 foaf:based near geo:2635167mdb:film/2014 movie:relatedBook bm:0743424425mdb:film/2014 movie:language lexvo:iso639-3/engmdb:director/8476 movie:director name “Stanley Kubrick”mdb:film/2685 movie:director mdb:director/8476mdb:film/2685 rdfs:label “A Clockwork Orange”mdb:film/424 movie:director mdb:director/8476mdb:film/424 rdfs:label “Spartacus”mdb:actor/29704 movie:actor name “Jack Nicholson”mdb:film/1267 movie:actor mdb:actor/29704mdb:film/1267 rdfs:label “The Last Tycoon”mdb:film/3418 movie:actor mdb:actor/29704mdb:film/3418 rdfs:label “The Passenger”geo:2635167 gn:name “United Kingdom”geo:2635167 gn:population 62348447geo:2635167 gn:wikipediaArticle wp:United Kingdombm:books/0743424425 dc:creator bm:persons/Stephen+Kingbm:books/0743424425 rev:rating 4.7bm:books/0743424425 scom:hasOffer bm:offers/0743424425amazonOfferlexvo:iso639-3/eng rdfs:label “English”lexvo:iso639-3/eng lvont:usedIn lexvo:iso3166/CAlexvo:iso639-3/eng lvont:usesScript lexvo:script/Latn

URI Literal

URI


RDF Graph

mdb:film/2014

“1980-05-23”



bm:books/0743424425

4.7

rev:rating


geo:2635167


gn:name

62348447

gn:population

mdb:actor/29704


movie:actor name

mdb:film/3418

“The Passenger”

refs:label

mdb:film/1267


refs:label

mdb:director/8476


movie:director name

mdb:film/2685


refs:label

mdb:film/424

“Spartacus”

refs:label

mdb:actor/30013

movie:relatedBook

scam:hasOffer






RDF Query Model – SPARQL

Query Model - SPARQL Protocol and RDF Query LanguageGiven U (set of URIs), L (set of literals), and V (set of variables), aSPARQL expression is defined recursively:

an atomic triple pattern, which is an element of

(U ∪ V )× (U ∪ V )× (U ∪ V ∪ L)

?x rdfs:label “The Shining”

P FILTER R, where P is a graph pattern expression and R is a built-inSPARQL condition (i.e., analogous to a SQL predicate)

?x rev:rating ?p FILTER(?p > 3.0)

P1 AND/OPT/UNION P2, where P1 and P2 are graph patternexpressions

Example:SELECT ?nameWHERE {

?m r d f s : l a b e l ?name . ?m movie : d i r e c t o r ?d .?d movie : d i r e c t o r n a m e ” S t a n l e y K u b r i c k ” .?m movie : r e l a t e d B o o k ?b . ?b r e v : r a t i n g ? r .FILTER(? r > 4 . 0 )

}© M. Tamer Ozsu Croucher ASI (2015/12/16-18) 48 / 96

SPARQL Queries

SELECT ?nameWHERE {


}

?m ?dmovie:director

?name

rdfs:label

?b

movie:relatedBook


movie:director name

?rrev:rating

FILTER(?r > 4.0)


Outline





Naıve Triple Store Design

SELECT ?nameWHERE {


}Subject Property Objectmdb:film/2014 rdfs:label “The Shining”mdb:film/2014 movie:initial release date “1980-05-23”mdb:film/2014 movie:director mdb:director/8476mdb:film/2014 movie:actor mdb:actor/29704mdb:film/2014 movie:actor mdb: actor/30013mdb:film/2014 movie:music contributor mdb: music contributor/4110mdb:film/2014 foaf:based near geo:2635167mdb:film/2014 movie:relatedBook bm:0743424425mdb:film/2014 movie:language lexvo:iso639-3/engmdb:director/8476 movie:director name “Stanley Kubrick”mdb:film/2685 movie:director mdb:director/8476mdb:film/2685 rdfs:label “A Clockwork Orange”mdb:film/424 movie:director mdb:director/8476mdb:film/424 rdfs:label “Spartacus”mdb:actor/29704 movie:actor name “Jack Nicholson”mdb:film/1267 movie:actor mdb:actor/29704mdb:film/1267 rdfs:label “The Last Tycoon”mdb:film/3418 movie:actor mdb:actor/29704mdb:film/3418 rdfs:label “The Passenger”geo:2635167 gn:name “United Kingdom”geo:2635167 gn:population 62348447geo:2635167 gn:wikipediaArticle wp:United Kingdombm:books/0743424425 dc:creator bm:persons/Stephen+Kingbm:books/0743424425 rev:rating 4.7bm:books/0743424425 scom:hasOffer bm:offers/0743424425amazonOfferlexvo:iso639-3/eng rdfs:label “English”lexvo:iso639-3/eng lvont:usedIn lexvo:iso3166/CAlexvo:iso639-3/eng lvont:usesScript lexvo:script/Latn

Easy to implementbut

too many self-joins!



SELECT ?nameWHERE {



SELECT T1 . o b j e c tFROM T as T1 , T as T2 , T as T3 ,

T as T4 , T as T5WHERE T1 . p=” r d f s : l a b e l ”AND T2 . p=” movie : r e l a t e d B o o k ”AND T3 . p=” movie : d i r e c t o r ”AND T4 . p=” r e v : r a t i n g ”AND T5 . p=” movie : d i r e c t o r n a m e ”AND T1 . s=T2 . sAND T1 . s=T3 . sAND T2 . o=T4 . sAND T3 . o=T5 . sAND T4 . o > 4 . 0AND T5 . o=” S t a n l e y K u b r i c k ”





SELECT ?nameWHERE {



SELECT T1 . o b j e c tFROM T as T1 , T as T2 , T as T3 ,

T as T4 , T as T5WHERE T1 . p=” r d f s : l a b e l ”AND T2 . p=” movie : r e l a t e d B o o k ”AND T3 . p=” movie : d i r e c t o r ”AND T4 . p=” r e v : r a t i n g ”AND T5 . p=” movie : d i r e c t o r n a m e ”AND T1 . s=T2 . sAND T1 . s=T3 . sAND T2 . o=T4 . sAND T3 . o=T5 . sAND T4 . o > 4 . 0AND T5 . o=” S t a n l e y K u b r i c k ”




Exhaustive Indexing

RDF-3X [Neumann and Weikum, 2008, 2009], Hexastore [Weisset al., 2008]

Strings are mapped to ids using a mapping table

Triples are indexed in a clustered B+ tree in lexicographic order

Create indexes for permutations of the three columns: SPO, SOP,PSO, POS, OPS, OSP

Original triple tableSubject Property Objectmdb: film/2014 rdfs:label “The Shining”mdb:film/2014 movie:initial release date “1980-05-23”mdb:director/8476 movie:director name “Stanley Kubrick”mdb:film/2685 movie:director mdb:director/8476

Encoded triple tableSubject Property Object

0 1 20 3 45 6 78 9 5

Mapping tableID Value0 mdb: film/20141 rdfs:label2 “The Shining”3 movie:initial release date4 “1980-05-23”5 mdb:director/84766 movie:director name7 “Stanley Kubrick”8 mdb:film/26859 movie:director


Exhaustive Indexing





Subject Property Object0 1 2

0 3 4

5 6 7

8 9 5...

......

B+ treeEasy queryingthrough mappingtable


Exhaustive Indexing






0 3 4

5 6 7

8 9 5...

......

B+ treeEasy queryingthrough mappingtable


Exhaustive Indexing–Query Execution

Each triple pattern can be answered by a range query

Joins between triple patterns computed using merge join

Join order is easy due to extensive indexing


0 3 4

5 6 7

8 9 5...

......

ID Value0 mdb: film/2014

1 rdfs:label

2 “The Shining”

3 movie:initial release date

4 “1980-05-23”

5 mdb:director/8476

6 movie:director name

7 “Stanley Kubrick”

8 mdb:film/2685

9 movie:director







0 3 4

5 6 7

8 9 5...

......


1 rdfs:label

2 “The Shining”


4 “1980-05-23”

5 mdb:director/8476



8 mdb:film/2685

9 movie:director

Advantages

I Eliminates some of the joins – they become range queries

I Merge join is easy and fast







0 3 4

5 6 7

8 9 5...

......


1 rdfs:label

2 “The Shining”


4 “1980-05-23”

5 mdb:director/8476



8 mdb:film/2685

9 movie:director

Advantages

I Eliminates some of the joins – they become range queries

I Merge join is easy and fast

Disadvantages

I Space usage


Property Tables

Grouping by entities; Jena [Wilkinson, 2006], DB2-RDF [Borneaet al., 2013]

Clustered property table: group together the properties that tend tooccur in the same (or similar) subjects

Property-class table: cluster the subjects with the same type ofproperty into one property table

Subject Property Objectmdb:film/2014 rdfs:label “The Shining”mdb:film/2014 movie:director mdb:director/8476mdb:film/2685 movie:director mdb:director/8476mdb:film/2685 rdfs:label “A Clockwork Orange”mdb:actor/29704 movie:actor name “Jack Nicholson”. . . . . . . . .

Subject refs:label movie:directormob:film/2014 “The Shining” mob:director/8476mob:film/2685 “The Clockwork Orange” mob:director/8476

Subject movie:actor namemdb:actor “Jack Nicholson”


Property Tables







Advantages

I Fewer joins

I If the data is structured, we have a relational system – similar tonormalized relations


Property Tables







Advantages

I Fewer joins

I If the data is structured, we have a relational system – similar tonormalized relations

Disadvantages

I Potentially a lot of NULLs

I Clustering is not trivial

I Multi-valued properties are complicated


Binary Tables

Grouping by properties: For each property, build a two-column table,containing both subject and object, ordered by subjects [Abadi et al.,2007, 2009]

Also called vertical partitioned tables

n two column tables (n is the number of unique properties in the data)


Subject Objectmdb:film/2014 mdb:director/8476mdb:film/2685 mdb:director/8476

movie:director

Subject Objectmob:film/2014 “The Shining”mob:film/2685 “The Clockwork Orange”

refs:label

Subject Objectmdb:actor/29704 “Jack Nicholson”

movie:actor name


Binary Tables






movie:director


refs:label


movie:actor name

Advantages

I Supports multi-valued properties

I No NULLs

I No clustering

I Read only needed attributes (i.e. less I/O)

I Good performance for subject-subject joins


Binary Tables






movie:director


refs:label


movie:actor name

Advantages

I Supports multi-valued properties

I No NULLs

I No clustering

I Read only needed attributes (i.e. less I/O)

I Good performance for subject-subject joins

Disadvantages

I Not useful for subject-object joins

I Expensive inserts


Graph-based Approach

Answering SPARQL query ≡ subgraph matching using homomorphism

gStore [Zou et al., 2011, 2014], chameleon-db [Aluc et al., 2013]

?m ?dmovie:director

?name

rdfs:label

?b

movie:relatedBook


movie:director name

?rrev:rating

FILTER(?r > 4.0)

mdb:film/2014

“1980-05-23”



bm:books/0743424425

4.7

rev:rating


geo:2635167


gn:name

62348447

gn:population

mdb:actor/29704


movie:actor name

mdb:film/3418

“The Passenger”

refs:label

mdb:film/1267


refs:label

mdb:director/8476


movie:director name

mdb:film/2685


refs:label

mdb:film/424

“Spartacus”

refs:label

mdb:actor/30013

movie:relatedBook

scam:hasOffer





SubgraphM

atching





?m ?dmovie:director

?name

rdfs:label

?b

movie:relatedBook


movie:director name

?rrev:rating

FILTER(?r > 4.0)

mdb:film/2014

“1980-05-23”



bm:books/0743424425

4.7

rev:rating


geo:2635167


gn:name

62348447

gn:population

mdb:actor/29704


movie:actor name

mdb:film/3418

“The Passenger”

refs:label

mdb:film/1267


refs:label

mdb:director/8476


movie:director name

mdb:film/2685


refs:label

mdb:film/424

“Spartacus”

refs:label

mdb:actor/30013

movie:relatedBook

scam:hasOffer





SubgraphM

atching

Advantages

I Maintains the graph structure

I Full set of queries can be handled





?m ?dmovie:director

?name

rdfs:label

?b

movie:relatedBook


movie:director name

?rrev:rating

FILTER(?r > 4.0)

mdb:film/2014

“1980-05-23”



bm:books/0743424425

4.7

rev:rating


geo:2635167


gn:name

62348447

gn:population

mdb:actor/29704


movie:actor name

mdb:film/3418

“The Passenger”

refs:label

mdb:film/1267


refs:label

mdb:director/8476


movie:director name

mdb:film/2685


refs:label

mdb:film/424

“Spartacus”

refs:label

mdb:actor/30013

movie:relatedBook

scam:hasOffer





SubgraphM

atching

Advantages

I Maintains the graph structure

I Full set of queries can be handled

Disadvantages

I Graph pattern matching is expensive


Two Systems

gStore

mdb:film/2014

bm:books/0743424425

mdb:director/8476

mdb:film/424mdb:film/2685

mdb:actor/29804

mdb:film/3418 mdb:film/1267

mdb:actor/30013

movie:ac

tor

moive:director

“Spartacus”moive:director moive:director

“Jack_Nicholson”


rdfs:label

“1980-05-23”

rdfs:label

moive:actor_name

y:hasBudget

y:has_box_office

“22000000#dollar”

movie:relatedBook4.7

rev:rating

bm:offers/0743424425

scam:hasOffer

y:hasBudget y:hasBudget

“21000000#dollar” “26589355#dollar” “12000000#dollar” “60000000#dollar”

y:has_box_office

movie:actor


“The Passager” “The last Tycoon”rdfs:label rdfs:label

“Scatman Crothers”

movie:initial_release_datemoive:actor_name

Fig. 2. An RDF graph G

?x

?y

?z

mdb:movierdf:type

moive:director

“*Jack*”moive:actor_name

y:hasBudget

?budget<30000000Desc, top10

movie:actor

SELECT ?x ?y WHERE{ ?x hasBudget ?budget. ?x rdf:type mdb:movie. ?x movie:director ?y. ?y movie:actor_name ?z. FILTER( regex(str(?z),``Jack'') AND (?budget <30000000) )}ORDER BY ?budgetLIMIT 10

Fig. 3. SPARQL and Query Graph Q

a query signature graph Q⇤, the encoding strategy is analogueto encoding RDF graphs.

The online query evaluation process consists of two steps:filtering and joining. First, we generate the candidates for eachquery node using VS⇤-tree. Then, applying a depth-first searchstrategy, we perform the multi-way join over these candidatelists to find the subgraph matches of SPARQL query Q overRDF graph G.

III. Techniques

In this section, we briefly discuss the techniques used ingStore system; full details are given in elsewhere [5], [6]. Ac-cording to our framework in Section II, we solve the SPARQLquery processing by subgraph matching over the signaturegraph. A key issue is that the proposed encoding and pruningstrategies should support, in a uniform manner, di↵erent kindsof data (such as strings and numeric data), and SPARQLqueries with di↵erent operators . We discuss the encoding andpruning methods in Section III-A. Another technical issue isthe index structure, which is discussed in Section III-B. Wealso present some system-oriented optimization, such as indexcaching strategy and multicore-based query optimization inour system.

A. Encoding Techniques

In gSore, answering SPARQL queries is equivalent tofinding subgraph matches of query graph Q over RDF graphG. If vertex v (in query Q) can match vertex u (in RDF graphG), each neighbor vertex and each adjacent edge of v shouldmatch to some neighbor vertex and some adjacent edge of u.Thus, given a vertex u in G, we encode each of its adjacentedge labels and the corresponding neighbor vertex labels into

System Architecture

Offline Online

Storage

Input Input

RDF Parser

RDF Graph Builder

Encoding Module

VS*-tree builder

RDF data

RDF Triples

RDF Graph

Signature Graph

Key-Value Store

VS*-treeStore

SPARQL Parser

SPARQL Query

Encoding Module

VS*-tree

Query Graph

Filter Module

Join Module

Signature Graph

Node Candidate

Results

Fig. 4. System Architecture

bitstrings, denoted as vS ig(u). We encode query Q with thesame encoding method. Consequently, the match between Qand G can be verified by simply checking the match betweencorresponding encoded bitstrings.

Given a vertex u, we encode each of its adjacent edgese(eLabel, nLabel) into a bitstring, where eLabel is the edgelabel and nLabel is the vertex label. This bitstring is callededge signature (i.e., eS ig(e)). It has two parts: eS ig(e).e,eS ig(e).n. The first part eS ig(e).e (M bits) denotes the edgelabel (i.e., eLabel) and the second part eS ig(e).n (N bits)denotes the neighbor vertex label (i.e., nLabel). The code ofvS ig(u) is formed by performing OR operator over all eS ig(e).Figure 5 illustrates the process.

mdb:film/2014

mdb:director/8476

mdb:actor/29804 moive:director

“1980-05-23”

y:hasBudget

“22000000#dollar”

movie:initial_release_date

movie:actor

e1 rdfs:label "The Shining"e2 movie:initial_release_date "1980-05-23"e3 movie:director mdb:director/8476e4 movie:actor mdb:actor/29704e5 movie:actor mdb:actor/30013e6 y:hasDuration 7140.0$#se7 y:hasBudget 22000000$#$e8 y:hasImdb "0081505"rdfs:label"The Shining"

hasDuration

hasDuration

"0081505"

y:hasImdb

eSig.e eSig.ne1 001000010 000010000101000e2 000110000 000000011100000e3 100100000 000010010000001e4 000010010 001001000000001e5 000010010 001001010000000e6 101000000 000001001100000e7 001010000 000010000001001e8 100010000 001000001001000

nSig 101110010 001011011101001

Fig. 5. Encoding Technique

1) Computing eS ig(e).e: Given an RDF repository, let |P|denote the number of di↵erent properties. If |P| is small, weset |eS ig(e).e| = |P|, where |eS ig(e).e| denotes the length ofthe bitstring, and build a 1-to-1 mapping between the propertyand the bit position. If |P| is large, we resort to the hashingtechnique. Let |eS ig(e).e| = M. Using an appropriate hashfunction, we set m out of M bits in eS ig(e).e to be ‘1’.

chameleon-db

Structural Index

...

Vertex Index

Spill Index

Clu

ster

Ind

ex

Sto

rag

eS

yst

em Sto

rag

eA

dvis

or

QueryEngine Plan Generation Evaluation


gStore

General Approach:

Work directly on the RDF graph and the SPARQL query graph

Use a signature-based encoding of each entity and class vertex tospeed up matching

Filter-and-evaluate

Use a false positive algorithm to prune nodes and obtain a set ofcandidates; then do more detailed evaluation on those

Use an index (VS∗-tree) over the data signature graph (has lightmaintenance load) for efficient pruning


1. Encode Q and G to Get Signature Graphs

Query signature graph Q∗

0100 0000 1000 000000010

0000 010010000

Data signature graph G∗

0010 1000

0100 0001

00001

1000 000100010

0000 0100

10000

0000 1000

10000

0000 0010

10000

0000 1001

00100

0001 000101000

0100 1000

01000

1001 1000

01000

0001 0100

01000


2. Filter-and-Evaluate

Query signature graph Q∗

0100 0000 1000 000000010

0000 010010000

Data signature graph G∗

0010 1000

0100 0001

00001

1000 000100010

0000 0100

10000

0000 1000

10000

0000 0010

10000

0000 1001

00100

0001 000101000

0100 1000

01000

1001 1000

01000

0001 0100

01000

Find matches of Q∗ oversignature graph G ∗

Verify each match inRDF graph G


How to Generate Candidate List

Two step process:1 For each node of Q∗ get lists of nodes in G∗ that include that node.2 Do a multi-way join to get the candidate list

Alternatives:

Sequential scan of G∗

Both steps are inefficient

Use S-trees

Height-balanced tree over signaturesRun an inclusion query for each node of Q∗ and get lists of nodes inG∗ that include that node.

• Given query signature q and a set of data signatures S , find alldata signatures si ∈ S where q&si = q

Does not support second step – expensive

VS-tree (and VS∗-tree)

Multi-resolution summary graph based on S-treeSupports both steps efficientlyGrouping by vertices




Alternatives:

Sequential scan of G∗


Use S-trees









Alternatives:Sequential scan of G∗


Use S-trees











Use S-trees











Use S-trees







S-tree Solution

1111 1111

0110 1111 1101 1101

0000 1110 0110 1001 1100 1001 1001 1101

0000 1000

0000 0100 0000 0010

0010 1000

0100 0001

1000 0001

0000 1001

0100 1000

1001 1000

0001 0100

0001 0001

005

004 006

001

002

003

007

011

008

009

010

d11

d21 d2

2

d31 d3

2 d33 d3

4

G 3

G 2

G 1

1000 00000100 000000010

0000 010010000

Possibly large join space!


S-tree Solution

1111 1111

0110 1111 1101 1101

0000 1110 0110 1001 1100 1001 1001 1101

0000 1000

0000 0100 0000 0010

0010 1000

0100 0001

1000 0001

0000 1001

0100 1000

1001 1000

0001 0100

0001 0001

005

004 006

001

002

003

007

011

008

009

010

d11

d21 d2

2

d31 d3

2 d33 d3

4

G 3

G 2

G 1

1000 00000100 000000010

0000 010010000



S-tree Solution

1111 1111

0110 1111 1101 1101

0000 1110 0110 1001 1100 1001 1001 1101

0000 1000

0000 0100 0000 0010

0010 1000

0100 0001

1000 0001

0000 1001

0100 1000

1001 1000

0001 0100

0001 0001

005

004 006

001

002

003

007

011

008

009

010

d11

d21 d2

2

d31 d3

2 d33 d3

4

G 3

G 2

G 1

1000 00000100 000000010

0000 010010000 002

011



S-tree Solution

1111 1111

0110 1111 1101 1101

0000 1110 0110 1001 1100 1001 1001 1101

0000 1000

0000 0100 0000 0010

0010 1000

0100 0001

1000 0001

0000 1001

0100 1000

1001 1000

0001 0100

0001 0001

005

004 006

001

002

003

007

011

008

009

010

d11

d21 d2

2

d31 d3

2 d33 d3

4

G 3

G 2

G 1

1000 00000100 000000010

0000 010010000 002

011

003

008



S-tree Solution

1111 1111

0110 1111 1101 1101

0000 1110 0110 1001 1100 1001 1001 1101

0000 1000

0000 0100 0000 0010

0010 1000

0100 0001

1000 0001

0000 1001

0100 1000

1001 1000

0001 0100

0001 0001

005

004 006

001

002

003

007

011

008

009

010

d11

d21 d2

2

d31 d3

2 d33 d3

4

G 3

G 2

G 1

1000 00000100 000000010

0000 010010000 002

011

003

008

004

009



S-tree Solution

1111 1111

0110 1111 1101 1101

0000 1110 0110 1001 1100 1001 1001 1101

0000 1000

0000 0100 0000 0010

0010 1000

0100 0001

1000 0001

0000 1001

0100 1000

1001 1000

0001 0100

0001 0001

005

004 006

001

002

003

007

011

008

009

010

d11

d21 d2

2

d31 d3

2 d33 d3

4

G 3

G 2

G 1

1000 00000100 000000010

0000 010010000 002

011

003

008

004

009on on



S-tree Solution

1111 1111

0110 1111 1101 1101

0000 1110 0110 1001 1100 1001 1001 1101

0000 1000

0000 0100 0000 0010

0010 1000

0100 0001

1000 0001

0000 1001

0100 1000

1001 1000

0001 0100

0001 0001

005

004 006

001

002

003

007

011

008

009

010

d11

d21 d2

2

d31 d3

2 d33 d3

4

G 3

G 2

G 1

1000 00000100 000000010

0000 010010000 002

011

003

008

004

009on on



VS-tree

1111 1111

0110 1111 1101 1101

0000 1110 0110 1001 1100 1001 1001 1101

0000 1000

0000 0100 0000 0010

0010 1000

0100 0001

1000 0001

0000 1001

0100 1000

1001 1000

0001 0100

0001 0001

005

004 006

001

002

003

007

011

008

009

010

d11

d21 d2

2

d31 d3

2 d33 d3

4

G 3

G 2

G 1

11101

1001010001 01100

10000 00001 01100

00010

10000

01000

01000

10000

10000

10000

1000000010

00100

01000

01000

01000

01000

Super edge


Pruning with VS-Tree

1111 1111

0110 1111 1101 1101

0000 1110 0110 1001 1100 1001 1001 1101

0000 1000

0000 0100 0000 0010

0010 1000

0100 0001

1000 0001

0000 1001

0100 1000

1001 1000

0001 0100

0001 0001

005

004 006

001

002

003

007

011

008

009

010

d11

d21 d2

2

d31 d3

2 d33 d3

4

G 3

G 2

G 1

11101

1001010001 01100

10000 00001 01100

00010

10000

01000

01000

10000

10000

10000

1000000010

00100

01000

01000

01000

01000

1000 00000100 000000010

0000 010010000

d32

d33

d33

d34

d31

d34

G 3

00010 10000

01000

003

008

002

011

004

009onon

Reduced join space!



1111 1111

0110 1111 1101 1101

0000 1110 0110 1001 1100 1001 1001 1101

0000 1000

0000 0100 0000 0010

0010 1000

0100 0001

1000 0001

0000 1001

0100 1000

1001 1000

0001 0100

0001 0001

005

004 006

001

002

003

007

011

008

009

010

d11

d21 d2

2

d31 d3

2 d33 d3

4

G 3

G 2

G 1

11101

1001010001 01100

10000 00001 01100

00010

10000

01000

01000

10000

10000

10000

1000000010

00100

01000

01000

01000

01000

1000 00000100 000000010

0000 010010000

d32

d33

d33

d34

d31

d34

G 3

00010 10000

01000

003

008

002

011

004

009onon

Reduced join space!



1111 1111

0110 1111 1101 1101

0000 1110 0110 1001 1100 1001 1001 1101

0000 1000

0000 0100 0000 0010

0010 1000

0100 0001

1000 0001

0000 1001

0100 1000

1001 1000

0001 0100

0001 0001

005

004 006

001

002

003

007

011

008

009

010

d11

d21 d2

2

d31 d3

2 d33 d3

4

G 3

G 2

G 1

11101

1001010001 01100

10000 00001 01100

00010

10000

01000

01000

10000

10000

10000

1000000010

00100

01000

01000

01000

01000

1000 00000100 000000010

0000 010010000

d32

d33

d33

d34

d31

d34

G 3

00010 10000

01000

003

008

002

011

004

009onon

Reduced join space!



1111 1111

0110 1111 1101 1101

0000 1110 0110 1001 1100 1001 1001 1101

0000 1000

0000 0100 0000 0010

0010 1000

0100 0001

1000 0001

0000 1001

0100 1000

1001 1000

0001 0100

0001 0001

005

004 006

001

002

003

007

011

008

009

010

d11

d21 d2

2

d31 d3

2 d33 d3

4

G 3

G 2

G 1

11101

1001010001 01100

10000 00001 01100

00010

10000

01000

01000

10000

10000

10000

1000000010

00100

01000

01000

01000

01000

1000 00000100 000000010

0000 010010000

d32

d33

d33

d34

d31

d34

G 3

00010 10000

01000

003

008

002

011

004

009onon

Reduced join space!



1111 1111

0110 1111 1101 1101

0000 1110 0110 1001 1100 1001 1001 1101

0000 1000

0000 0100 0000 0010

0010 1000

0100 0001

1000 0001

0000 1001

0100 1000

1001 1000

0001 0100

0001 0001

005

004 006

001

002

003

007

011

008

009

010

d11

d21 d2

2

d31 d3

2 d33 d3

4

G 3

G 2

G 1

11101

1001010001 01100

10000 00001 01100

00010

10000

01000

01000

10000

10000

10000

1000000010

00100

01000

01000

01000

01000

1000 00000100 000000010

0000 010010000

d32

d33

d33

d34

d31

d34

G 3

00010 10000

01000

003

008

002

011

004

009onon

Reduced join space!



1111 1111

0110 1111 1101 1101

0000 1110 0110 1001 1100 1001 1001 1101

0000 1000

0000 0100 0000 0010

0010 1000

0100 0001

1000 0001

0000 1001

0100 1000

1001 1000

0001 0100

0001 0001

005

004 006

001

002

003

007

011

008

009

010

d11

d21 d2

2

d31 d3

2 d33 d3

4

G 3

G 2

G 1

11101

1001010001 01100

10000 00001 01100

00010

10000

01000

01000

10000

10000

10000

1000000010

00100

01000

01000

01000

01000

1000 00000100 000000010

0000 010010000

d32

d33

d33

d34

d31

d34

G 3

00010 10000

01000

003

008

002

011

004

009onon

Reduced join space!


Adaptivity to Workload

Applications that rely on RDF data are increasingly popular and aremore varied [Verborgh et al., 2014]

Data that are being handled are far more heterogeneous [Duan et al.,2011]

SPARQL queries are becoming more diverse [Arias et al., 2011] anddynamic [Kirchberg et al., 2011]

An experiment [Aluc et al., 2014a]

No single system is a sole winner across all queriesNo single system is the sole loser across all queries, eitherThere can be 2–5 orders of magnitude difference in the performance(i.e., query execution time) between the best and the worst system fora given queryThe winner in one query may timeout in anotherPerformance difference widens as dataset size increases








Can existing systems cope with these trends – workload diversity &dynamism

No! [Aluc et al., 2014b]

I Fragmented data

I Suboptimal pruning by indexes

I Unnecessarily large sets of intermediate result tuples








Our proposal: Idea behind chameleon-db

I When designing and implementing an RDF data management system,assume nothing about the workload upfront

I Organize data dynamically and purely based on the workload


Group-by-Query Approach

v1

v21

A

v21

v20

B

v1

v98

A

v98

v30

C

v1

v250

A

v250

v40

D

v0

v32

A

v0

v52 C

v0

v80

C

v0

v66C

v0

v47

B

v6

v7

A

v6

v8

B

v6

v9

C

C1 C2 C3 C4 C5

Characteristics:

Records are not necessarily of fixed length

Records are not grouped into tables

Records do not necessarily share the same set of RDF predicates

Each record represents a very tiny part of the RDF graph



v1

v21

A

v21

v20

B

v1

v98

A

v98

v30

C

v1

v250

A

v250

v40

D

v0

v32

A

v0

v52 C

v0

v80

C

v0

v66C

v0

v47

B

v6

v7

A

v6

v8

B

v6

v9

C

C1 C2 C3 C4 C5

?x ?yA

?y ?z?b

Figure: Query 1



v1

v21

A

v21

v20

B

v1

v98

A

v98

v30

C

v1

v250

A

v250

v40

D

v0

v32

A

v0

v52 C

v0

v80

C

v0

v66C

v0

v47

B

v6

v7

A

v6

v8

B

v6

v9

C

C1 C2 C3 C4 C5

?x ?yA

?y ?z?b

Figure: Query 1



v1

v21

A

v21

v20

B

v1

v98

A

v98

v30

C

v1

v250

A

v250

v40

D

v0

v32

A

v0

v52 C

v0

v80

C

v0

v66C

v0

v47

B

v6

v7

A

v6

v8

B

v6

v9

C

C1 C2 C3 C4 C5

?x ?yA

?y ?z?b

Figure: Query 1



v1

v21

A

v21

v20

B

v1

v98

A

v98

v30

C

v1

v250

A

v250

v40

D

v0

v32

A

v0

v52 C

v0

v80

C

v0

v66C

v0

v47

B

v6

v7

A

v6

v8

B

v6

v9

C

C1 C2 C3 C4 C5

?a

?z

C

?a

?x

A

?a

?y

B

Figure: Query 2



v1

v21

A

v21

v20

B

v1

v98

A

v98

v30

C

v1

v250

A

v250

v40

D

v0

v32

A

v0

v52 C

v0

v80

C

v0

v66C

v0

v47

B

v6

v7

A

v6

v8

B

v6

v9

C

C1 C2 C3 C4 C5

?a

?z

C

?a

?x

A

?a

?y

B

Figure: Query 2



v1

v21

A

v21

v20

B

v1

v98

A

v98

v30

C

v1

v250

A

v250

v40

D

v0

v32

A

v0

v52 C

v0

v80

C

v0

v66C

v0

v47

B

v6

v7

A

v6

v8

B

v6

v9

C

C1 C2 C3 C4 C5

?a

?z

C

?a

?x

A

?a

?y

B

Figure: Query 2



v1

v21

A

v21

v20

B

v1

v98

A

v98

v30

C

v1

v250

A

v250

v40

D

v0

v32

A

v0

v52 C

v0

v80

C

v0

v66C

v0

v47

B

v6

v7

A

v6

v8

B

v6

v9

C

C1 C2 C3 C4 C5

?a

?z

C

?a

?x

A

?a

?y

B

Figure: Query 2

Advantages

I Data are physically clustered for the workload

I Better pruning by the indexes

I Fewer intermediate result tuples


Challenges

Physical Data Layout: As the workloads change, the way data aregrouped together may no longer be suitable

Hierarchical Clustering Algorithm [Aluc et al., 2015]Tunable-LSH [Aluc et al., 2015]

Indexing: Indexing upfront is not a choice

Query Evaluation: Can we execute queries efficiently even when thephysical layout is constantly changing? [Aluc et al., 2015]


chameleon-db

Prototype system [Aluc et al., 2013]35,000 lines of code in C++ under Linux (plus code for SPARQL 1.0parser)

Structural Index

...

Vertex Index

Spill Index

Clu

ster

Inde

xS

tora

geS

yste

m Sto

rage

Adv

isor

QueryEngine Plan Generation Evaluation


Outline





Remember the Environment

Distributed environment

Some of the data sites canprocess SPARQL queries –SPARQL endpoints

Not all data sites can processqueries

Alternatives

Data re-distribution + querydecompositionData re-distribution + partialevaluationSPARQL federation: justprocess at SPARQL endpointsLive querying (next section)






Alternatives







Alternatives

Data re-distribution + querydecomposition

Data re-distribution + partialevaluationSPARQL federation: justprocess at SPARQL endpointsLive querying (next section)






Alternatives

Data re-distribution + querydecompositionData re-distribution + partialevaluation

SPARQL federation: justprocess at SPARQL endpointsLive querying (next section)






Alternatives

Data re-distribution + querydecompositionData re-distribution + partialevaluationSPARQL federation: justprocess at SPARQL endpoints

Live querying (next section)






Alternatives



Distributed RDF Processing [Kaoudi and Manolescu, 2015]

RDF data warehouse is partitioned and distributedRDF data D = {D1, . . . ,Dn}Allocate each Di to a site

Partitioning alternativesTable-based (e.g., [Husain et al., 2011])Graph-based (e.g., [Huang et al., 2011; Zhang et al., 2013])Unit-based (e.g., [Gurajada et al., 2014; Lee and Liu, 2013])

SPARQL query decomposed Q = {Q1, . . . ,Qk}Distributed execution of {Q1, . . . ,Qk} over {D1, . . . ,Dn}

I High performance

I Great for parallelizing centralized RDF data

I May not be possible to re-partition and re-allocate Web data (i.e.,LOD)

I Query decomposition may not be easy






I High performance









I High performance









I High performance





Distributed RDF Processing – 2

Data summary-based approaches

Build summaries (index) for the distributed RDF datasets (e.g., [Atreet al., 2010; Prasser et al., 2012])

SPARQL query Q = {Q1, . . . ,Qk}Distributed execution of {Q1, . . . ,Qk} using the data summary

I No data re-partitioning and re-allocation

I Have to scan the data at each site

I Index over distributed data with maintenance concerns


















SPARQL Endpoint Federation

Consider only the SPARQL endpoints for query execution

No data re-partitioning/re-distribution

Consider D = D1 ∪ D2 ∪ . . . ∪ Dn; Di : SPARQL endpoint

SPARQL query decomposed Q = {Q1, . . . ,Qk}Distributed execution of {Q1, . . . ,Qk} over {D1, . . . ,Dn}Systems

DARQ, FedX [Schwarte et al., 2011], SPLENDID [Gorlitz and Staab,2011], ANAPSID [Acosta et al., 2011]

I Data integration approach

I May be the only way to proceed if data is distributed

I Not all RDF data storage points are SPARQL endpoints







Systems


























Distributed SPARQL Using Partial Query Evaluation

Two steps:1 Evaluate a query at each site to find local matches

Query is the function and each Di is the known inputInner match or local partial match

2 Assemble the partial matches to get final resultCrossing matchCentralized assemblyDistributed assembly

D1

D2

D3

D4

Crossing match


Distributed SPARQL Using Partial Query Evaluation

Two steps:1 Evaluate a query at each site to find local matches

Query is the function and each Di is the known inputInner match or local partial match

2 Assemble the partial matches to get final resultCrossing matchCentralized assemblyDistributed assembly

D1

D2

D3

D4

Crossing match


Outline





Linked Object Data – Closer Look


Globally Distributed Network of Data


Traditional Hypertext-based Web Access

IMDb WorldBook

Data exposedto the Webvia HTML


Linked Data Publishing Principles

IMDb WorldBook

(http://...linkedmdb.../Shining,releaseDate, 23 May 1980)(http://...linkedmdb.../Shining, filmLocation, http://cia.../UK)(http://...linkedmdb.../29704,actedIn, http://...linkedmdb.../Shining)

...

(http://cia.../UK, hasPopulation, 63230000)...

Shi

ning

UK

Data model: RDFGlobal identifier: URIAccess mechanism: HTTPConnection: data links


Live Query Processing

Not all data resides at SPARQLendpoints

Freshness of access to dataimportant

Potentially countably infinitedata sources

Live querying

On-line executionOnly rely on linked dataprinciples

Alternatives

Traversal-based approachesIndex-based approachesHybrid approaches


Linked Data Model [Hartig, 2012]

Web of Linked Data

Given a finite or countably infinite set D of Linked Documents, a Web ofLinked Data is a tuple W = (D, adoc, data) where:

I D ⊆ D,

I adoc is a partial mapping from URIs to D, and

I data is a total mapping from D to finite sets of RDF triples.


Linked Data Model [Hartig, 2012]

Web of Linked Data

Given a finite or countably infinite set D of Linked Documents, a Web ofLinked Data is a tuple W = (D, adoc, data) where:

I D ⊆ D,

I adoc is a partial mapping from URIs to D, and

I data is a total mapping from D to finite sets of RDF triples.

Data Links

A Web of Linked Data W = (D, adoc, data)contains a data link from document d ∈ D todocument d ′ ∈ D if there exists a URI u suchthat:

I u is mentioned in an RDF triplet ∈ data(d), and

I d ′ = adoc(u).© M. Tamer Ozsu Croucher ASI (2015/12/16-18) 81 / 96

SPARQL Query Semantics in Live Querying

Full-web semantics

Scope of evaluating a SPARQL expression is all Linked DataQuery result completeness cannot be guaranteed by any (terminating)execution

Reachability-based query semantics

Query consists of a SPARQL expression, a set of seed URIs S , and areachability condition cScope: all data along paths of data links that satisfy the conditionComputationally feasible


SPARQL Query Semantics in Live Querying

Full-web semantics

Scope of evaluating a SPARQL expression is all Linked DataQuery result completeness cannot be guaranteed by any (terminating)execution

Reachability-based query semantics

Query consists of a SPARQL expression, a set of seed URIs S , and areachability condition cScope: all data along paths of data links that satisfy the conditionComputationally feasible


Traversal Approaches

Discover relevant URIs recursively bytraversing (specific) data links at queryexecution runtime [Hartig, 2013;Ladwig and Tran, 2011]

Implements reachability-based querysemantics

Start from a set of seed URIsRecursively follow and discover newURIs

Important issue is selection of seed URIs

Retrieved data serves to discover newURIs and to construct result








Advantages

Easy to implement.No data structure to maintain.








Advantages

Easy to implement.No data structure to maintain.

Disadvantages

Possibilities for parallelized data retrieval are limitedRepeated data retrieval introduces significant query latency.


Index Approaches

Use pre-populated index to determine relevant URIs (and to avoid asmany irrelevant ones as possible)

Different index keys possible; e.g., triple patterns [Umbrich et al.,2011]

Index entries a set of URIsIndexed URIs may appear multiple times (i.e., associated with multipleindex keys)Each URI in such an entry may be paired with a cardinality (utilized forsource ranking)

Key: tp Entry: {uri1, uri2, , urin}

GET urii


Index Approaches





GET urii

Advantages

Data retrieval can be fully parallelizedReduces the impact of data retrieval on query execution time


Index Approaches





GET urii

Advantages

Data retrieval can be fully parallelizedReduces the impact of data retrieval on query execution time

Disadvantages

Querying can only start after index constructionDepends on what has been selected for the indexFreshness may be an issueIndex maintenance


Hybrid Approach

Perform a traversal-based execution using a prioritized list of URIs tolook up [Ladwig and Tran, 2010]

Initial seed from the pre-populated index

Non-seed URIs are ranked by a function based on information in theindex

New discovered URIs that are not in the index are ranked accordingto number of referring documents


Acknowledgements

This presentation draws upon collaborative research and discussions withthe following colleagues (in alphabetical order)

Gunes Aluc, U. Waterloo

Khaled Ammar, U. Waterloo

Khuzaima Daudjee, U. Waterloo

Young Han, U. Waterloo

Olaf Hartig, U. Waterloo

Lei Chen, Hong Kong UST

Lei Zou, Peking Univ.


Thank you!

Research supported by



References I

Abadi, D. J., Marcus, A., Madden, S., and Hollenbach, K. (2009). SW-Store: a verticallypartitioned DBMS for semantic web data management. VLDB J., 18(2):385–406.

Abadi, D. J., Marcus, A., Madden, S. R., and Hollenbach, K. (2007). Scalable semanticweb data management using vertical partitioning. In Proc. 33rd Int. Conf. on VeryLarge Data Bases, pages 411–422.

Acosta, M., Vidal, M.-E., Lampo, T., Castillo, J., and Ruckhaus, E. (2011). ANAPSID:an adaptive query processing engine for SPARQL endpoints. In Proc. 10th Int.Semantic Web Conf., pages 18–34.

Aluc, G., Ozsu, M. T., and Daudjee, K. (2015). Clustering RDF databases usingTunable-LSH. CoRR, abs/1504.02523.

Aluc, G., Hartig, O., Ozsu, M. T., and Daudjee, K. (2014a). Diversified stress testing ofRDF data management systems. In Proc. 13th Int. Semantic Web Conf., pages197–212.

Aluc, G., Ozsu, M. T., and Daudjee, K. (2014b). Workload matters: Why RDFdatabases need a new design. Proc. VLDB Endowment, 7(10):837–840.


References II

Aluc, G., Ozsu, M. T., Daudjee, K., and Hartig, O. (2013). chameleon-db: aworkload-aware robust RDF data management system. Technical Report CS-2013-10,University of Waterloo. Available at https://cs.uwaterloo.ca/sites/ca.computer-science/files/uploads/files/CS-2013-10.pdf.

Aluc, G., Ozsu, M. T., Daudjee, K., and Hartig, O. (2015). Executing queries overschemaless RDF databases. In Proc. 31st Int. Conf. on Data Engineering, pages807–818.

Ammar, K. and Ozsu, M. T. (2015). Approaches to graph processing – an overview. Inpreparation.

Arias, M., Fernandez, J. D., Martınez-Prieto, M. A., and de la Fuente, P. (2011). Anempirical study of real-world SPARQL queries. CoRR, abs/1103.5043.

Atre, M., Chaoji, V., Zaki, M. J., and Hendler, J. A. (2010). Matrix “bit” loaded: Ascalable lightweight join query processor for rdf data. In Proc. 19th Int. World WideWeb Conf., pages 41–50.

Bornea, M. A., Dolby, J., Kementsietsidis, A., Srinivas, K., Dantressangle, P., Udrea, O.,and Bhattacharjee, B. (2013). Building an efficient RDF store over a relationaldatabase. In Proc. ACM SIGMOD Int. Conf. on Management of Data, pages121–132.


https://cs.uwaterloo.ca/sites/ca.computer-science/files/uploads/files/CS-2013-10.pdf

https://cs.uwaterloo.ca/sites/ca.computer-science/files/uploads/files/CS-2013-10.pdf

References III

Bu, Y., Howe, B., Balazinska, M., and Ernst, M. D. (2012). The HaLoop approach tolarge-scale iterative data analysis. VLDB J., 21(2):169–190.

Dean, J. and Ghemawat, S. (2008). Mapreduce: Simplified data processing on largeclusters. Commun. ACM, 51(1):107–113.

Duan, S., Kementsietsidis, A., Srinivas, K., and Udrea, O. (2011). Apples and oranges:a comparison of RDF benchmarks and real RDF datasets. In Proc. ACM SIGMODInt. Conf. on Management of Data, pages 145–156.

Gonzalez, J. E., Xin, R. S., Dave, A., Crankshaw, D., Franklin, M. J., and Stoica, I.(2014). GraphX: graph processing in a distributed dataflow framework. In Proc. 11thUSENIX Symp. on Operating System Design and Implementation, pages 599–613.

Gorlitz, O. and Staab, S. (2011). SPLENDID: SPARQL endpoint federation exploitingVOID descriptions. In Proc. 2nd Int. Workshop on Consuming Linked Data.

Gurajada, S., Seufert, S., Miliaraki, I., and Theobald, M. (2014). TriAD: A distributedshared-nothing RDF engine based on asynchronous message passing. In Proc. ACMSIGMOD Int. Conf. on Management of Data, pages 289–300.

Han, M., Daudjee, K., Ammar, K., Ozsu, M. T., Wang, X., and Jin, T. (2014). Anexperimental comparison of Pregel-like graph processing systems. Proc. VLDBEndowment, 7(12):1047–1058.


References IV

Hartig, O. (2012). SPARQL for a web of linked data: Semantics and computability. InProc. 9th Extended Semantic Web Conf., pages 8–23.

Hartig, O. (2013). SQUIN: a traversal based query execution system for the web oflinked data. In Proc. ACM SIGMOD Int. Conf. on Management of Data, pages1081–1084.

Huang, J., Abadi, D. J., and Ren, K. (2011). Scalable SPARQL querying of large RDFgraphs. Proc. VLDB Endowment, 4(11):1123–1134.

Husain, M. F., McGlothlin, J., Masud, M. M., Khan, L. R., and Thuraisingham, B.(2011). Heuristics-based query processing for large RDF graphs using cloudcomputing. IEEE Trans. Knowl. and Data Eng., 23(9):1312–1327.

Kaoudi, Z. and Manolescu, I. (2015). RDF in the clouds: A survey. VLDB J., 24:67–91.

Kirchberg, M., Ko, R. K. L., and Lee, B.-S. (2011). From linked data to relevant data –time is the essence. CoRR, abs/1103.5046.

Ladwig, G. and Tran, T. (2010). Linked data query processing strategies. In Proc. 9thInt. Semantic Web Conf., pages 453–469.

Ladwig, G. and Tran, T. (2011). SIHJoin: Querying remote and local linked data. InProc. 8th Extended Semantic Web Conf., pages 139–153.


References V

Lee, K. and Liu, L. (2013). Scaling queries over big rdf graphs with semantic hashpartitioning. Proc. VLDB Endowment, 6(14):1894–1905.

Li, F., Ooi, B. C., Ozsu, M. T., and Wu, S. (2014). Distributed data management usingMapReduce. ACM Comput. Surv., 46(3):Article No. 31.

Low, Y., Gonzalez, J., Kyrola, A., Bickson, D., Guestrin, C., and Hellerstein, J. M.(2012). Distributed graphlab: A framework for machine learning in the cloud. Proc.VLDB Endowment, 5(8):716–727.

Malewicz, G., Austern, M. H., Bik, A. J. C., Dehnert, J. C., Horn, I., Leiser, N., andCzajkowski, G. (2010). Pregel: a system for large-scale graph processing. In Proc.ACM SIGMOD Int. Conf. on Management of Data, pages 135–146.

Michiardi, P. (2015). Introduction to spark internals. Slideshare. Available from:http://www.slideshare.net/michiard/introduction-to-spark-internals?

qid=511145e7-79d7-41d8-a133-9e705d4933c3&v=qf1&b=&from_search=11 [Lastretrieved: 9 July 2015].

Neumann, T. and Weikum, G. (2008). RDF-3X: a RISC-style engine for RDF. Proc.VLDB Endowment, 1(1):647–659.


http://www.slideshare.net/michiard/introduction-to-spark-internals?qid=511145e7-79d7-41d8-a133-9e705d4933c3&v=qf1&b=&from_search=11

http://www.slideshare.net/michiard/introduction-to-spark-internals?qid=511145e7-79d7-41d8-a133-9e705d4933c3&v=qf1&b=&from_search=11

References VI

Neumann, T. and Weikum, G. (2009). The RDF-3X engine for scalable management ofRDF data. VLDB J., 19(1):91–113.

Prasser, F., Kemper, A., and Kuhn, K. A. (2012). Efficient distributed query processingfor autonomous rdf databases. In Proc. 15th Int. Conf. on Extending DatabaseTechnology, pages 372–383.

Schwarte, A., Haase, P., Hose, K., Schenkel, R., and Schmidt, M. (2011). Fedx: Afederation layer for distributed query processing on linked open data. In Proc. 8thExtended Semantic Web Conf., pages 481–486.

Tian, Y., Balmin, A., Corsten, S. A., Tatikonda, S., and McPherson, J. (2013). From“think like a vertex” to “think like a graph”. Proc. VLDB Endowment, 7(3):193–204.

Umbrich, J., Hose, K., Karnstedt, M., Harth, A., and Polleres, A. (2011). Comparingdata summaries for processing live queries over linked data. World Wide Web J.,14(5-6):495–544.

Verborgh, R., Hartig, O., Meester, B. D., Haesendonck, G., Vocht, L. D., Sande, M. V.,Cyganiak, R., Colpaert, P., Mannens, E., and de Walle, R. V. (2014). Queryingdatasets on the web with high availability. In Proc. 13th Int. Semantic Web Conf.,pages 180–196.


References VII

Weiss, C., Karras, P., and Bernstein, A. (2008). Hexastore: sextuple indexing forsemantic web data management. Proc. VLDB Endowment, 1(1):1008–1019.

Wilkinson, K. (2006). Jena property table implementation. Technical ReportHPL-2006-140, HP Laboratories Palo Alto.

Yan, D., Cheng, J., Lu, Y., and Ng, W. (2014). Blogel: A block-centric framework fordistributed computation on real-world graphs. Proc. VLDB Endowment,7(14):1981–1992.

Zaharia, M. (2016). An Architecture for Fast and General Data Processing on LargeClusters. ACM Books. Forthcoming.

Zaharia, M., Chowdhury, M., Das, T., Dave, A., Ma, J., McCauley, M., Franklin, M. J.,Shenker, S., and Stoica, I. (2012). Resilient distributed datasets: A fault-tolerantabstraction for in-memory cluster computing. In Proc. 9th USENIX Symp. onNetworked Systems Design & Implementation, pages 2–2.

Zaharia, M., Chowdhury, M., Franklin, M. J., Shenker, S., and Stoica, I. (2010). Spark:Cluster computing with working sets. In Proc. 2nd USENIX Workshop on Hot Topicsin Cloud Computing, pages 10–10.


References VIII

Zhang, X., Chen, L., Tong, Y., and Wang, M. (2013). EAGRE: Towards scalable I/Oefficient SPARQL query evaluation on the cloud. In Proc. 29th Int. Conf. on DataEngineering, pages 565–576.

Zou, L., Mo, J., Chen, L., Ozsu, M. T., and Zhao, D. (2011). gStore: answeringSPARQL queries via subgraph matching. Proc. VLDB Endowment, 4(8):482–493.

Zou, L., Ozsu, M. T., Chen, L., Shen, X., Huang, R., and Zhao, D. (2014). gStore: Agraph-based SPARQL query engine. VLDB J., 23(4):565–590.


Data & Analytics

Graph lecture