You are on page 1of 17

BenchmarkingGr

aph
Analyt
icSys
tems:
Ti
ger
Graph,
Neo4j
,andTi
tan
Execut
iveSummar
y:Ast
hewor
ld’
sfir
stNat
ivePar
all
elGr
aph,
theTi
ger
GraphTM sy
stem
l
oads,
stor
es,
andquer
iesdat
afast
ert
hanbot
hNeo4j
andTi
tan.

Ti
ger
Graphhas4xt
o100xf
ast
ergr
apht
raver
salandquer
yresponset
imes
compar
edt
o bot
hNeo4jandTi
tan,
whenusi
ngasi
ngl
emachi
ne.

Wi
thTi
ger
Gr quer
aph, yspeedi
ncr
easesal
mostl
inear
lywi
tht
henumberof

machi
nes.Ev
enonPageRank
,requi
ri
ngmassi
venode-
to-
nodecommuni
cat
ion,
Ti
ger
Graphst
il
lachi
eves4.
7xspeedupwi
th8xmachi
nes.
Ti
ger
Graph’
sinher
entpar
all
el
ism enabl
esi
ttol
oaddat
aonl
ine50xf
ast
ert
hanNeo4j
and25xf
ast
ert
hanTi
tan,ev
enwhenusi
ngasi
ngl
emachi
ne.

Ti
ger
Graph’
shi
ghcompr
essi
onr
educesst
or oabout1/
aget 5thatofbot
hNeo4j

andTi
tan.

Thi
sbenchmar
kst
udyex
ami
nest
hedat
aloadi
ngandquer
yper
for
manceofTi
ger
Graph,

Neo4j
,andTi
tan.
Benchmar
ktopi
csi
ncl
udet
hef
oll
owi
ng:
Dat
aLoadi
ng

Loadi
ngcapabi
l
iti
es
Loadi
ngt
imeandspeed
St
oragesi
zeofl
oadeddat
a
Quer
yPer
for
mance

Quer
yresponset
imes-f
orK-
Nei
ghbor
hood

Quer
yresponset
imes-f
orWeak
lyConnect
edComponent
s

Scal
abi
l
ityofquer
yresponset
imes

© 2017Ti
ger
Graph.
All
Right
sReser
ved. 1
1.Benchmar
kSet
up

Gr
aphdat
abase/
anal
yti
cssyst
emst
est
ed:

Ti
ger
Graph(
for
mer
lyGr
aphSQL)0.
8

Neo4j
3.1.
3Communi
tyEdi
ti
on

Ti
tanGr
aph1.
0.0

Har
dwar
epl
atf
orm

Thi
sbenchmar
kusesanAmaz
onEC2i
nst
ancet
ype.I
tisl
argeenoughf
ort
het
est

dat
aset
sandbal
ancesst
rongcomput
ati
onal
capabi
l
itywi
theconomy
.

Vi
rt
ual
machi
ne:
Amaz
onEC2,
C3.
8xl
argei
nst
ancet
ype

32v
CPUs,
equi
val
entt
o108ECU

60Gi
Bmemor
yand2x320GBSSDdi
sk

Addi
ti
onal
stor
age:
1TBAmaz
onEBS

Tabl
e1-Dat
aset
s

NAME DESCRI
PTI
ONANDSOURCE VERTI
CES EDGES

gr
aph500-
22 Sy
nthet
icgr
aph 2.
4M 64M
ht
tp:
//gr
aph500.
org

t
wit
ter
_rv
.net Twi
tt
eruser
-fol
l
owerdi
rect
edgr
aph 41.
6M 1.
47B
ht
tp:
//an.
kai
st.
ac.
kr/
traces/
WWW2010.
html

Foreachgr
aph,
ther
awdat
aar
efor
mat
tedasasi
ngl
etab-
separ
atededgel
i
st:

U1 U3

U1 U4

U2 U3

2
Ver
ti
cesdonothav
eat
tr
ibut
essot
her
eisnoneedf
orasepar
atev
ert
exl
i
st.

Dat
aLoadi
ngTest
s

Thedat
aloadi
ngt
est
sexami
nedt
heset
hreear
eas:

Loadi
ngcapabi
l
iti
es

Loadi
ngt
imeandspeed

St
oragesi
zeofl
oadeddat
a

LOADI
NGCAPABI
LITI
ES
Bef
oreconsi
der
ingper
for
mance,
wefir
stcompar
efunct
ional
capabi
l
iti
es.
Thi
sisnecessar
y

t
olear
nhowt
oopt
imi
zet
hel
oadi
ngf
oreachgr
aphdat
apl
atf
orm.

Ti
ger
Graphcomposesal
oadi
ngj
obi
nit
snat
iveGSQLl
anguaget
her
ebycr
eat
ingaRESTf
ul

endpoi
nt.
Thej
obcant
henber
unei
thert
hrought
heGSQLshel
lorbysubmi
tt
inganHTTP

r
equestt
otheRESTf
ulser
ver
.Al
ll
oadi
ngt
akespl
aceonl
i
newhi
l
eot
herdat
aaccess

ser
vicesar
erunni
ng.
TheRESTf
uldi
rectr
equesti
susedf
ort
hist
est
.

Neo4jof
fer
stwol
oader
s:of
fl
inel
oadi
ngusi
ngneo4j
-i
mpor
t1
andonl
i
nel
oadi
ngwi
tht
he

phercommand2.
LOADCSVCy Thei
rcapabi
l
iti
esar
equi
tedi
ff
erent
.Bot
hmet
hodswer
e

t
est
edf
ort
hisBenchmar
k.

Ti
tanpr
ovi
dessy
stem set
ti
ngsandadv
icef
orper
for
mingbul
kloadi
ng3,
buti
tdoesnot

of
feranat
ivefil
e-t
o-DBl
oadi
ngl
anguage.
WeusedCassandr
athr
if
tast
hest
orageback
end

andwr
oteacust
om l
oadi
ngpr
ocedur
einJav
ausi
ngt
heTi
nker
popGr
aphSt
ruct
ureAPI
.

ht
1
tps:
//neo4j
.
com/
docs/
oper
ati
ons-
manual
/cur
rent
/tut
ori
al
/i
mpor
t-t
ool
/
ht
2
tp:
//neo4j
.com/docs/
devel
oper-
manual/curr
ent/
cypher/
clauses/
load-
csv
/
ht
3
t
p:/
/s3.t
hinkaur
eli
us.
com/docs/
tit
an/
0.5.4/
bulk
-loadi
ng.
html

3
Tabl
e2-Compar
isonofl
oadi
ngcapabi
li
ti
es

FEATURE TI
GERGRAPH NEO4J-
IMPORT NEO4J-
CYPHER TI
TAN

Bui
lt
-i
nloadi
ng
l
anguage YES YES YES NO

Bulkonl
inel
oadi
ng
( o100B+edges) YES
100M t NO NO YES

I
ncrement
aldat
a
YES NO YES YES
l
oading

I
ndexbui
l
tdur
ing
YES NO YES YES
l
oadi
ng

Vert
exID
YES NO YES NO
dedupl
i
cati
on

Ti
ger
Graphof
fer
sthemostconv
eni
entl
oadi
ngenvi
ronment
,wi
thbul
konl
ine
l
oadi
ng,
abui
lt
-inl
oadi
ngl
anguage,
andbui
lt
-ini
ndexi
ng.

LOADI
NGTI
MEANDSPEED
Togetgoodper
for
mancei
nNeo4j
-i
mpor
t,aschemai
ndexmustbecr
eat
edsepar
atel
y.The

t
otal
loadi
ngt
imef
orNeo4j
incl
udesbot
hindexcr
eat
ionanddat
aloadi
ng.
Thet
imef
or

cr
eat
ingt
hei
ndexal
onei
sshowni
npar
ent
hesesi
nTabl
e3.
Ast
hel
oadi
ngpr
ogr
essed,

Neo4j
Cypher
’sl
oadi
nggotsl
owerandsl
owerasi
tcheck
edf
ordupl
i
cat
ever
texI
Ds.
Neo4j

Cy
pherdi
dnotcompl
etet
hel
oadi
ngt
estwi
thi
n24hour
s,sot
hel
oadi
ngwashal
tedand

l
oadi
ngonasmal
l
erdat
asetwasat
tempt
ed.

4
Tabl
e3-Dat
aset
s

NAME TI
GERGRAPH NEO4J-
OFFLI
NE NEO4J-
CYPHER TI
TAN

gr
aph500-
22 139s (
13s)116s Di
dnotcompl
ete 3002s
i
n24hour
s

t
wit
ter
_rv
.net 1815s (
135s)3739s Di
dnotcompl
ete 44654s
i
n24hour
s

Thechar
tbel
owpr
esent
sthesel
oadi
ngt
imes.

Fi
gur
e1-Nor
mal
iz
eddat
aloadi
ngt
imes

Ti
ger
Graphi
s2xf
ast
ert
hanNeo4j
-of
fl
ineand25xf
ast
ert
hanTi
tanatl
oadi
ng
t
hel
argedat
asett
wit
ter
-rv
.

Byr
educi
ngt
hedat
asetsi
zet
o1mi
l
li
onl
i
nes,
Neo4j
Cypher
’sl
oadi
ngspeed(
numberof

dat
ali
nes/
second)becamemeasur
abl
e.Per
iodi
cbat
chcommi
t,
wit
hbat
chsi
ze=100,
000,

© 2017Ti
ger
Graph.
All
Right
sReser
ved. 5
wasusedt
oopt
imi
zeNeo4j
'
sonl
i
nel
oadi
ngper
for
mance.
Tabl
e4i
ncl
udest
hedat
a

l
oadi
ngspeedr
esul
tsofbot
hTi
ger
GraphandNeo4j
Cypherusi
ngt
hissmal
l
erdat
aset
.

Tabl
e4-Dat
aloadi
ngspeed

DATASET TI
GERGRAPH NEO4JCYPHER

t
wit
ter
-rv(
first1M l
i
nes) 809Kl
i
nes/
sec 14Kl
i
nes/
sec

Thechar
tbel
owpr
esent
sthesel
oadi
ngt
imes.

Fi
gur
e2-Av
erageLoadi
ngSpeed

Ti
ger
Graph l
oads1M edgesmor
ethan50xf
ast
ert
hanNeo4jCypher
.
Neo4jCypherdi
dnotl
oadev
ent
hesmal
lestdat
aset(
64M edges)
.

STORAGESI
ZEOFLOADEDDATA
Thesi
zeoft
hel
oadeddat
aisani
mpor
tantconsi
der
ati
on,
especi
al
lyf
ori
n-memor
y

© 2017Ti
ger
Graph.
All
Right
sReser
ved. 6
RAW DATA TI
GERGRAPH NEO4J TI
TAN

967MB 454MB 2,
369MB 2,
515MB

24,
375MB 11,
038MB 51,
047MB 50,
028MB

Ti
ger
Graphst
oresl
argedat
aset
sinappr
oxi
mat
ely1/
5thespaceusedbybot
h
Neo4jandTi
tan.

7
DATALOADI
NGSUMMARY
Onl
yTi
ger
Graphcandoonl
inedat
aloadi
ngf
orl
arge-
scal
edat
a.

Ti
ger
Graph’
shi
ghcompr
essi
onr
educesst
oraget
oabout1/
5thatofbot
h
Neo4jandTi
tan.

3.Quer
yPer
for
manceTest
s

TheQuer
yper
for
mancet
est
sexami
nedt
heset
hreear
eas:

Quer
yresponset
imes-f
orK-
Nei
ghbor
hood

Quer
yresponset
imes-f
orWeak
lyConnect
edComponent
s

Scal
abi
l
ityofquer
yresponset
imes

QUERYRESPONSETI
MES-FORK-
NEI
GHBORHOOD
Thek
-nei
ghbor
hoodquer
y,whi
chask
sforal
lthev
ert
iceswhi
char
ewi
thi
nkhopsofa

st
art
ingv
ert
ex,
isagoodmeasur
eofgr
apht
rav
ersal
per
for
mance.
Theone-
nei
ghbor
hood

ofav
ert
exi
ssi
mpl
yit
ssetofadj
acentv
ert
ices.
Wecomput
etheav
eragequer
yti
mef
or

1000r
andoml
ysel
ect
edst
art
ingv
ert
icesi
neachdat
aset
.Eachst
art
ingv
ert
exhasa

di
ff
erentnei
ghbor
hoodsi
zeN;
wer
epor
ttheav
erageN.
Quer
yex
ecut
iont
imei
smeasur
ed

f
rom t
hecl
i
entsi
de(
cli
ent
srunont
hesamemachi
neast
hedat
abase)
,whi
chi
ncl
udest
he

net
wor
ktr
ansf
ert
ime.
Tof
ocusont
hegr
apht
rav
ersal
andt
omi
nimi
zet
henet
wor
kout
put

t
ime,
weout
putonl
ythesi
zeoft
hek
-nei
ghbor
hood,
inst
eadoft
hecompl
etel
i
stofv
ert
ices.

© 2017Ti
ger
Graph.
All
Right
sReser
ved. 8
Bot
hTi
ger
GraphandNeo4j
cani
mpl
ementt
hek
-nei
ghbor
hoodquer
yef
fici
ent
ly
.

Ti
ger
Graphal
l
owst
heusert
oadj
ustt
henumberoft
hreadsandi
ntheset
est
s16t
hreads

wer
eused.
Tit
anusest
heGr
eml
i
nquer
ylanguageandt
heonl
yreasonabl
ewaywef
ound

t
oimpl
ementk
-nei
ghbor
hoodwasbyusi
ngt
hesubgr
aphAPI
.Thet
est
srev
eal
thatt
he

subgr
aphAPIi
sver
yresour
ce-
heav
y,mak
ingi
tver
ydi
fficul
ttor
unt
he2-
nei
ghbor
hood

quer
ytocompl
eti
on.
Inst
ead,
Tit
anper
for
msnot
iceabl
ybet
terwhenk
-nei
ghbor
hoodi
s

comput
edusi
ng t
hesetofv
ert
icest
hatar
eex
act
lykst
epsf
rom t
hesour
ce.
Thebel
ow

t
abl
esi
ncl
udet
hequer
yti
mesf
orTi
ger
Graph,
Neo4j
,Ti
tanusi
ngt
ruek
-nei
ghbor
hood,
and

Ti
tanusi
ngt
hek
-shel
lappr
oxi
mat
ion.

Tabl
e6-One-
Nei
ghbor
hoodquer
yti
me

DATASET AVGN TI
GERGRAPH NEO4J TITAN TI
TAN
(
K-NEIGH) (
K-SHELLAPPROX)

gr
aph500-
22 4082 3.
3ms 14.
4ms 3064.
9ms 14.
2ms

t
wit
ter
_rv
.net 12271 7.
5ms 55.
0ms 34.
1ms 39.
8ms

Tabl
e7-Two-
Nei
ghbor
hoodquer
yti
me

DATASET AVGN TI
GERGRAPH NEO4J TITAN TI
TAN
(
K-NEIGH) (
K-SHELLAPPROX)

gr
aph500-
22 457400 0.
19s 19.
0s Outofmem 11.
1s

t
wit
ter
_rv
.net 1747130 0.
77s 21.
2s 61.
6s 58.
7s

Theone-
nei
ghbor
hoodquer
yti
mef
orTi
tanf
ort
hegr
aph500-
22dat
asethasbeenomi
tt
ed

f
rom t
hechar
tbel
ow.
Ther
esponset
imewassohi
gh(
3064.
9ms)andf
aroutofr
angeof

t
heot
herv
aluest
hati
twoul
dmak
ethechar
tunr
eadabl
e.

© 2017Ti
ger
Graph.
All
Right
sReser
ved. 9
Fi
gur
e4-One-
nei
ghbor
hoodquer
yti
me

I
nthet
wo-
nei
ghbor
hoodcase,
Tit
anr
anoutofmemor
yont
hegr
aph500-
22dat
aset
.

Fi
gur
e5–Nor
mal
iz
edTwo-
nei
ghbor
hoodquer
yti
me

© 2017Ti
ger
Graph.
All
Right
sReser
ved. 10
Ti
ger
Graphi
s4xt
o7xf
ast
ert
hanNeo4jont
heone-
nei
ghbor
hoodquer
yand
27xt
o100xf
ast
ert
hanNeo4jont
het
wo-
nei
ghbor
hoodquer
y.

Ti
tani
s900xsl
owert
hanTi
ger
Graphont
heone-
nei
ghbor
hoodquer
y,and
coul
dnotfini
sht
het
wo-
nei
ghbor
hoodquer
y.

QUERYRESPONSETI
MES-FORWEAKL
YCONNECTEDCOMPONENTS
Aweak
lyconnect
edcomponent(
WCC)i
sthemax
imal
setofv
ert
icesandt
hei
rconnect
ing

edgeswhi
chcanr
eachoneanot
her
,i
fthedi
rect
ionofdi
rect
ededgesi
signor
ed.
TheWCC

quer
yfindsal
ltheWCCsi
nagr
aph.
Thi
squer
yrequi
rest
hatev
eryv
ert
exandev
eryedgebe

t
rav
ersed.

Wecompar
edTi
ger
Grapht
oNeo4j
.Ti
ger
Graph'
sGSQLquer
ylanguagei
sfl
exi
bleenough

t
oletust
rysev
eral
appr
oaches.
Wei
mpl
ement
edWCCt
woway
s.Fi
rst
,WCC-
global
mak
es

ev
eryv
ert
exast
art
ingpoi
ntandex
pandsf
rom eacht
ofindWWCsi
npar
all
el
;second,

WCC-
BFSr
unsabr
eadt
h-fir
stsear
chal
gor
it
hm f
rom anunl
abel
edv
ert
ex,
assi
gnsi
tanew

WCCi
d,andl
abel
sal
lreachabl
ever
ti
ceswi
tht
hesameWCCi
d.WCC-
global
hasbet
ter

par
all
el
ism,
whi
l
eWCC-
BFShasl
oweral
gor
it
hm compl
exi
ty
.Ti
ger
Graph'
sWCC-
BFSwas

about2xf
ast
ert
hanWCC-
global
,soonl
ytheWCC-
BFSt
imesar
erepor
ted.

ForNeo4j
,weuset
heNeo4j
nat
iveWCCi
mpl
ement
ati
onpr
ovi
dedi
ntheNeo4j
AOPC

pr
ocedur
es(
rel
ease3.
1.3.
6).
TheAPOC'
simpl
ement
ati
onofWCCadopt
sanal
gor
it
hm

si
mil
art
oWCC-
BFS.

© 2017Ti
ger
Graph.
All
Right
sReser
ved. 11
Tomi
nimi
zet
henet
wor
kov
erheadt
ime,
wedesi
gnedt
hequer
iest
oret
urnonl
ythenumber

ofweak
lyconnect
edcomponent
s,i
nst
eadofr
epor
ti
ngt
hef
ull
member
shi
pofeach

component
.

Tabl
e8-Weakl
yconnect
edcomponentquer
yti
me

DATASET TI
GERGRAPH-
BFS-
16 NEO4J

gr
aph500-
22 8.
7s 73.
0s

t
wit
ter
-rv 217.
8s Outofmemor
y

Ti
ger
Graphi
s8xf
ast
ert
hanNeo4jont
hesmal
lerdat
aset
.
Neo4jhasamuchl
argermemor
yfoot
pri
ntandr
anoutofmemor
ywhen
r
unni
ngont
hel
argert
wit
ter
-rvdat
aset
.

SCALABI
LITYOFQUERYRESPONSETI
MES
I
nourfinal
test
,weex
ami
nehowt
heNat
ivePar
all
elGr
aphar
chi
tect
ureofTi
ger
Graph

enabl
est
hequer
yresponset
imet
odecr
easeast
henumberofmachi
nesusedi
ncr
eases.

Fort
hist
est
,weusedasl
i
ght
lydi
ff
erentAmaz
onEC2i
nst
ancet
ype(
DB.
R3.
2xl
arge)
.We

usedt
het
wit
ter
_rvdat
asetandt
hewel
l
-knownPageRankquer
y(r
ankev
eryv
ert
ex)

r
epeat
ed10t
imes.
Wet
est
edwi
th1,
2,4,
and8machi
nesi
npar
all
el
.Thegr
aphwas

par
ti
ti
onedi
ntoequal
l
ysi
zedsegment
sacr
ossal
lthemachi
nesused.

© 2017Ti
ger
Graph.
All
Right
sReser
ved. 12
Tabl
e9-Scal
abi
li
tyofquer
yresponset
imes

NO.
OFMACHI
NES 1 2 3 4

av
eragequer
yti
me 81.
4s 51.
8s 28.
9s 17.
2s

Speedup 1.
0 1.
57 2.
81 4.
73

PageRanki
sani
ter
ati
veal
gor
it
hm whi
cht
rav
ersesev
eryedgedur
ingev
eryi
ter
ati
on.Thi
s

meanst
her
eismuchcommuni
cat
ionbet
weent
hemachi
nes,
wit
hinf
ormat
ionbei
ngsent

f
rom onepar
ti
ti
ont
oanot
her
.Despi
tet
hiscommuni
cat
ionov
erhead,
Tiger
Graph’
sNPG

ar
chi
tect
urest
il
lsucceedsi
nachi
evi
nga4.
7xspeedupwi
th8machi
nes.

Fi
gur
e6-Quer
yresponset
imevs.Numberofmachi
nes

Neo4j
isasi
ngl
eser
verandnotadi
str
ibut
edsy
stem.
ThusNeo4j
cannotpar
ti
ti
onagr
aph

acr
ossmul
ti
plemachi
nesandhandl
ebi
ggr
aphdat
a.Ti
ger
Graphi
sthefir
standonl
ytr
uly

di
str
ibut
ednat
ivepar
all
elgr
aphsy
stem.
Agr
aphi
nTi
ger
Graphcanbepar
ti
ti
onedand

© 2017Ti
ger
Graph.
All
Right
sReser
ved. 13
st
oredonmul
ti
plemachi
nes.

QUERYPERFORMANCESUMMARY
Ti
ger
Graphhas4xt
o100xf
ast
ergr
apht
rav
ersalandquer
yresponset
imes

compar
edt
obot
hNeo4jandTi
tan.
Wi
thTi
ger
Graph,
quer
yspeedi
ncr
easesasaddi
ti
onalmachi
nesar
eadded,

achi
evi
nga4.
7xspeedi
ncr
easewi
th8machi
nes.

© 2017Ti
ger
Graph.
All
Right
sReser
ved. 14
AboutTi
ger
Graph
Ti
ger
Graphi
sthewor
ld’
sfir
stReal
-Ti
meGr
aphAnal
yti
csPl
atf
orm power
edbyNat
ive
Par
all
elGr
aph(
NPG)t
echnol
ogy
.Ti
ger
Graphf
ulfil
l
sthet
ruepr
omi
seandbenefit
soft
he

gr
aphpl
atf
orm bysuppor
ti
ngr
eal
-t
imedeepl
i
nkanal
yti
csf
orent
erpr
iseswi
thcompl
ex
andcol
ossal
amount
sofdat
a.Ti
ger
Graph’
spr
ovent
echnol
ogyi
susedbycust
omer
s

i
ncl
udi
ngAl
i
pay
,VI
SA,
Sof
tBank
,St
ateGr
idCor
por
ati
onofChi
na,
WishandEl
ement
um.

FoundedbyYuXu,
Ph.
D.i
n2012,
Tiger
Graphi
sfundedbyQi
mingVC,
Bai
du,
AntFi
nanci
al
,

AMECl
oud,
Mor
adoVent
ures,
ZodNaz
em,
DanhuaCapi
tal
andDCVC.
Tiger
Graphi
sbased
i
nRedwoodCi
ty
,CA.
Lear
nmor
eatwww.
ti
ger
graph.
com.

© 2017Ti
ger
Graph.
All
Right
sReser
ved. 15
© 201
7Ti
ger
Graph

You might also like