1
E ec i eness o NoSQL and NewSQL
Da abases in Mobile Ne wo k E en Da a:
Cassand a and Pa S eam/Kine ic
Pe i Ko i an a, Ma ko Junkka i
Facul y o Na u al Sciences, Compu e Sciences, Uni e si y o Tampe e, FI-33014, Finland,
ko i an a.pe i.S@s uden .u a. i, ma ko.junkka i@u a. i
ABSTRACT
Con inuously g owing amoun o da a has inspi ed seeking mo e and mo e e icien da abase solu ions o s o ing
and manipula ing da a. In big da a se s, NoSQL da abases ha e been es ablished as al e na i es o adi ional
SQL da abases. The e ec i eness o hese da abases has been widely es ed, bu he es s ocused only on key- alue
da a ha is s uc u ally e y simple. Many applica ion domains, such as elecommunica ion, in ol e mo e complex
da a s uc u es. Huge amoun o Mobile Ne wo k E en (MNE) da a is p oduced by an inc easing numbe o
mobile and ubiqui ous applica ions. MNE da a is s uc u ally p ede e mined and ypically con ains a la ge numbe
o columns. Applica ions ha handle MNE da a a e usually inse in ensi e, as a huge amoun o da a a e
gene a ed du ing ush hou s. NoSQL p o ides high scalabili y and i s column amily s o es sui s MNE da a well,
bu NoSQL does no suppo ACID ea u es o he adi ional ela ional da abases. NewSQL is a new kind o
da abases, which p o ide he high scalabili y o NoSQL while s ill main aining ACID gua an ees o he adi ional
DBMS. In he pape , we e alua ion NEM da a s o ing and agg ega ing e iciency o Cassand a and
Pa S eam/Kine ic da abases and aim o ind ou whe he he new kind o da abase echnology can clea ly b ing
pe o mance ad an ages o e legacy da abase echnology and o e s an al e na i e o exis ing solu ions. Among
he column amily s o es o NoSQL, Cassand a is especially a good choice o inse in ensi e applica ions due o
i s way o handle da a inse ions. Pa S eam is a no el and ad anced NewSQL like da abase and is ecen ly
in eg a ed in o Cisco Kine ic. The esul s o he e alua ion show ha Pa S eam is much as e han Cassand a
when s o ing and agg ega ing MNE da a and he NewSQL is a e y s ong al e na i e o exis ing da abase
solu ions o inse in ensi e applica ions.
TYPE OF PAPER AND KEYWORDS
Sho Communica ion: pe o mance e alua ion, da a s o age, da a agg ega ion, inse ion, elecommunica ion,
NoSQL, NewSQL, Cassand a, Pa S eam, Kine ic, Mobile Ne wo k E en , MNE da a
1 INTRODUCTION
In elecommunica ion, e e bigge da a se s mus be
manipula ed because he numbe o ansac ions and he
amoun o da a associa ed wi h he ansac ions inc ease
cons an ly. In e na ional and na ional laws and s anda ds
de e mine wha kind o da a mus be s o ed abou a
single ansac ion, and hus hey de e mine he s uc u e
o da a. Due o he inc easing amoun o Mobile
Ne wo k E en (MNE) da a i is essen ial o in es iga e
possible solu ions o manipula e MNE da a. NoSQL
[13][14] and NewSQL [14][26] da abases a e mode n
Open Access
Open Jou nal o Da abases (OJDB)
Volume 5, Issue 1, 2018
www. onpub.com/ojdb
ISSN 2199-3459
© 2018 by he au ho s; licensee RonPub, Lübeck, Ge many. This a icle is an open access a icle dis ibu ed unde he e ms and condi ions
o he C ea i e Commons A ibu ion license (h p://c ea i ecommons.o g/licenses/by/4.0/).
Open Jou nal o Da abases (OJDB), Volume 5, Issue 1, 2018
2
solu ions o manipula e big da a se s. In his p esen
s udy, we e alua e and compa e he e iciency o
NoSQL da abases and NewSQL da abases in s o ing and
que ying MNE da a. Applica ions o MNE da a a e
inse in ensi e, and hus he main ocus o he
e alua ion is o compa e he s o ing speed o da abases.
Mo e gene ally, ou esea ch ques ion is whe he new
da abase solu ions b ing addi ional alue compa ed o
exis ing legacy SQL based solu ions.
The adi ional SQL da abases a e usually designed
o be ope a ed on one se e node. This is he way hey
can o e ACID (A omici y, Consis ency, Isola ion, and
Du abili y) p ope ies. Howe e , he d awback o his
ea u e is he lack o ho izon al scalabili y. Depending
on he implemen a ion, clus e ing is possible. Fo
example, in O acle i is possible o di ide he da abase
ables in o di e en se e nodes. Howe e , NoSQL
da abases enable o sp ead he da a a ound a clus e pe
da a ow based on he p ima y key. This is why he e has
been he demand o NoSQL da abases as hey o e
mo e ho izon al scalabili y. They also o e simple da a
models ha may be mo e e icien han SQL in ce ain
use cases. NewSQL is a class o SQL da abase sys ems,
which seek o achie e high pe o mance and scalabili y
o NoSQL while s ill gua an eeing he ACID p ope ies
o adi ional DBMS. By compa ing be ween NoSQL
da abases and NewSQL da abases, we aim o ind ou
whe he i is possible o combine he bes sides o he
SQL and NoSQL da abases and s ill pe o m well in ou
use case.
Among NoSQL da abases, column amily s o es and
documen s o es a e s uc u ally sui able o
manipula ing NME da a. We selec ed column amily
s o e Cassand a because i s inse ion speed is e icien
and i s da a iles ake less s o age space han JSON based
documen s o e implemen a ions. Among NewSQL
da abases we selec ed Pa S eam [7] because i is
sui able o managing MNE da a. Fo example, i
suppo s he geo-dis ibu ed da abase solu ion ha is
essen ial o MNE da a. As o he summe 2018, a e
Pa S eam was acqui ed by Cisco (www.cisco.com) ha
is a ma ke leade in he a eas o IT and ne wo k. Cisco
in eg a es i in o Cisco Kine ic sys em and does no o e
Pa S eam as a s and-alone p oduc any mo e [10], bu
Cisco s ill p o ides all documen s on Pa S eam online
[8]. The unc ionali y o Pa S eam is now a pa o
Cisco Kine ic, and hence we also e e o his da abase
by Pa S eam/Kine ic. In he es se ing we use o iginal
da abases because du ing ou es pe iods o MNE da a
we had only he license o Pa S eam bu no Kine ic.
MNE da a consis s o di e en epo s ha a e
gene a ed by elecommunica ions a ic. The mos
common epo is he RAB (Radio Access Bea e ) epo
ha is sen when a adio access bea e is c ea ed. A adio
access bea e p o ides a connec ion be ween a use
equipmen and a ne wo k se ice. I is c ea ed
p ac ically e e y ime when a use equipmen , o
example a mobile phone, ies o connec o a mobile
phone ne wo k h ough a base s a ion. These epo s
consis o s uc u ed da a ha ha e alues ep esen ing
se e al me ics om he base s a ion and in o ma ion
abou he use equipmen such as IMEI (In e na ional
Mobile Equipmen Iden i y) and IMSI (In e na ional
Mobile Subsc ibe Iden i y). These, in u n, con ain
coded in o ma ion on he coun y, ne wo ks and he
ou e h ough which a mobile phone plan has been
connec ed. Fo example, IMSI is a 64-bi ield ypically
ep esen ed as a 15 digi numbe whe e i s h ee digi s
de e mine a coun y, and a mobile ne wo k code and a
mobile subsc ip ion iden i ica ion numbe ollow.
Fo a mobile phone plan, he e is in o ma ion on he
coun y, ne wo ks and he ou e h ough which a mobile
phone plan has been connec ed. A connec ion in ol es
in o ma ion on di e en kinds o a ea codes and he
ou e h ough which he connec ion is o med. A
ne wo k s a ion ca ies i s own in o ma ion. All his
in o ma ion is collec ed in o a RAB epo ha is s o ed
in he con ex o a MNE e en . Depending on a e sion,
a RAB epo con ains abou 100 da a en ies, bu one
en y may con ain a alue ha ing di e en kind o coded
da a.
In a ypical scena io, a huge amoun o RAB epo s
mus be s o ed du ing a sho pe iod o ime. This
happens especially du ing ush hou s when many use
equipmen eques s RAB. The e o e, he s o ing speed
plays he mos essen ial ole in manipula ing RAB
epo s. Mos o he da a will no be u ilized, bu he
s o ing is necessa y o acing possible p oblems o
acking calls in se ious c iminal cases. The RAB epo s
can also be used analyzing he load o a ne wo k in a
speci ic a ea. The e o e, agg ega ion que ies a e
essen ial when analyzing he epo s. Fu he mo e,
pa e n ma ching que ies such as ‘like’ a e needed o
isola e pa s o he codes (e.g. he coun y code o IMSI).
In he e alua ion, we simula e eal wo ld mul i-
columned da a om he a ea o elecommunica ion, he
da a s o ing and agg ega ion pe o mance o he
NewSQL-like Pa S eam da abase and he NoSQL
Cassand a da abase a e e alua ed o e di e en amoun
o MNE da a using di e en numbe o h eads. The
esul s o he e alua ion show ha Pa S eam is
d ama ically as e han Cassand a in s o ing da a and i
bene i s om inc easing he numbe o h eads.
Pa S eam also ou pe o ms he adi ional SQL solu ion
in he inse ion speed o da a. The e iciency o
agg ega ion que ies depends on he column on which he
que y is ocused. I agg ega ion que ies do no ocus on
any speci ic columns, Pa S eam is no able as e han
Cassand a. We also conclude ha Cassand a does no
suppo pa e n ma ching que ies ha a e essen ial o
manipula ing NME da a.
In his s udy wo k, we compa e NoSQL and
NewSQL-like da abases whe eas exis ing s udies ocus
mainly on NoSQL da abases. To ou bes knowledge, no
P. Ko i an a, M. Junkka i: E ec i eness o NoSQL and NewSQL Da abases in Mobile Ne wo k E en Da a: Cassand a and Pa S eam/Kine ic
3
esea ch esul s ha e been published on e iciency
e alua ion o ei he Pa S eam o Kine ic and on he
pe o mance compa ison be ween NoSQL da abases
and NewSQL da abases.
The es o he pape is o ganized as ollows. In
Sec ion 2, we pe o m a li e a u e e iew on e iciency
s udies o NoSQL da abases. Sec ion 3 in es iga es
di e en da abase models in o de o ind sui able
da abases, which will be used in his s udy o e alua e
he pe o mance o da abases o MNE applica ions. In
Sec ion 4, we in oduce he con en o he Radio Access
Bea e (RAB) epo s. RAB epo s a e he Mobile
Ne wo k E en (MNE) da a and a e be used in he
e alua ion. Sec ion 5 desc ibes he gene a ion o RAB
da a and he e alua ion se ing. The esul s o e alua ion
a e p esen ed in Sec ion 6. Sec ion 7 discusses he
e alua ion esul s and in es iga es u he esea ch
ques ions. Finally, conclusions a e gi en in Sec ion 8.
2 RELATED WORK
The e has been a la ge amoun o esea ch on he
pe o mance o he NoSQL da abases [1][2][4][9][11]
[17][18][19][20][25]. Yahoo! Cloud Se ing Bench-
ma k [9][27] is he mos popula es ing en i onmen o
key- alue da a. Mo e complex da a a e used in
benchma king documen s o es wi h SQL da abases in
[23] and 0. Oli ei a and Be na dino [21] ha e compa ed
NewSQL da abases MemSQL and Vol DB using he
TPC-H es se ha is also a mo e complex da a se
con aining se e al ables and hei mu ual ela ionships.
We ocus on he s udies whe e column amily s o es a e
compa ed wi h o he da abases. In he ollowing, we
p esen he la es pe o mance es s whe e Cassand a is
compa ed wi h column amily s o es HBase,
Hype able, documen s o es MongoDB, Couchbase,
Ra enDB, CouchDB, key- alue da abases Ae ospike,
Redis, mul imodel da abase O ien DB and ela ional
da abase MS SQL Exp ess.
In he Da as ax s udy [11], Cassand a e sion 1.1.6,
HBase e sion 1.1.1 and MongoDB e sion 2.2.2 ha e
been compa ed. Yahoo! Cloud Se ing Benchma k was
used as a es ool. Load, ead, w i e and scan es s we e
made wi h di e en s ess le els and di e en amoun s
o clus e nodes. Read, inse , upda e and scan la ency
we e also es ed. Cassand a had clea ly he bes
pe o mance among he da abases. Especially, when he
amoun o clus e nodes was inc eased Cassand a was
much ahead lea ing HBase second and MongoDB hi d.
Nelubin and Engbe [20] compa ed Cassand a,
MongoDB, Couchbase and Ae ospike. In hei s udy,
he pe o mance o he da abases was compa ed using
Yahoo! Cloud Se ice Benchma king Tool. Da abases
we e compa ed o inse ion h oughpu , maximum
h oughpu and la encies in balanced wo kload (50%
w i e and 50% ead) and ead hea y wo kload (95% ead
and 5% upda e) in SSD (Solid S a e D i e)-backed and
in-memo y da ase s. The es s measu ed aw key- alue
pe o mance o he da abases. In hese es s, Ae ospike
and Couchbase had clea ly be e pe o mance
compa ed wi h Cassand a and MongoDB. Ae ospike
ou pe o med Couchbase in ead-hea y wo kloads and
Couchbase ou pe o med Ae ospike in balanced ead-
w i e wo kloads. One o he easons o he good
pe o mance o Ae ospike was ha i had been well
op imized o SSD disks ha we e used in his es . Bo h
Ae ospike and Couchbase a e designed o key- alue
based que ies and hese da abases we e expec ed o
pe o m be e han mo e complex Cassand a and
MongoDB. Howe e , pu e key- alue pe o mance is no
wha we a e looking o as MNE applica ions usually
equi e mo e complex que ies.
Li and Manoha an [19] compa ed MongoDB e sion
1.8.5, Ra enDB e sion 960, CouchDB e sion 1.2.0,
Cassand a e sion 1.1.2, Hype able e sion 0.9.6,
Couchbase e sion 1.8.0 and MS SQL Exp ess e sion
10.50.1600.1. The s udy es ed ins an ia ing a bucke o
key- alue pai s, eading alues behind keys, c ea ing
and upda ing key- alue pai s, dele ing key- alue pai s
and e ching all he keys. Ra enDB, Hype able and
MongoDB we e he as es whe eas CouchDB,
Couchbase and SQL Exp ess we e he slowes in
c ea ing he bucke . The ead pe o mance lis o
da abases om he as es o he slowes was as ollows:
Couchbase, MongoDB, SQL Exp ess, Hype able,
CouchDB, Cassand a and Ra enDB. Wi h w i e
pe o mance, he co esponding lis was Couchbase,
MongoDB, Cassand a, Hype able, SQL Exp ess,
Ra enDB and CouchDB, and wi h dele e pe o mance
Couchbase, MongoDB, SQL Exp ess, Cassand a,
Hype able, CouchDB and Ra enDB. In e ching all he
keys, he es obse a ion was ha all he da abases e ch
keys quickly excep CouchDB. SQL Exp ess was he
as es o doing his ope a ion.
One o he in e es ing indings o his s udy was ha
adi ional da abase MS SQL Exp ess pe o med be e
han some o he NoSQL da abases. Thus, al hough
NoSQL da abases should pe o m be e in key- alue
based que ies compa ed wi h adi ional da abases, hey
do no always pe o m be e han adi ional SQL
da abases. The e was only a small co ela ion be ween
pe o mance and da a models. Ra enDB and CouchDB
we e no good in ead, w i e and dele e ope a ions.
Couchbase and MongoDB we e o e all he as es o
ead, w i e and dele e ope a ions. Cassand a was slow in
ead ope a ions bu good in w i e and dele e ope a ions.
Anyway, Cassand a had he bes pe o mance among
column amily s o es.
Ab amo a and o he s [2] compa ed Cassand a
e sion 1.2.1, HBase e sion 0.94.10, MongoDB
e sion 2.4.6, O ien DB e sion 1.5 and Redis e sion
2.6.14. Among hese da abases O ien DB can be used as
Open Jou nal o Da abases (OJDB), Volume 5, Issue 1, 2018
4
Table 1: Top ou anking o NoSQL da abase pe o mance es s
Tes
1.
2.
3.
4.
Klein e al. [17]
Cassand a
Riak
MongoDB
-
Da as ax [11]
Cassand a
HBase
MongoDB
-
Nelubin & Engbe [20]
Couchbase
Ae ospike
MongoDB
Cassand a
Li & Manoha an [19]
Couchbase
MongoDB
Cassand a
Hype able
Ab amo a e al. [2]
Redis
Cassand a
HBase
MongoDB
a documen s o e and a g aph da abase. Da abases we e
es ed wi h Yahoo! Cloud Se ing Benchma k p og am.
Read and w i e ope a ions we e es ed wi h 600,000
eco ds. Tes s ocused on compa ing he execu ion
speed o ge and pu ope a ions wi h di e en wo kloads
o ead and upda e ope a ions. Redis was clea ly he
as es o he es ed da abases, Cassand a he second
as es , HBase hi d, MongoDB ou h. The slowes was
O ien DB when compa ing he o e all execu ion ime o
wo kloads. One o he easons o he poo pe o mance
o he O ien DB was ha i keeps eco ds in he disk
a he han loading hem in o memo y. O he eason
men ioned was ha O ien DB ook mo e esou ces han
wha was a ailable in he es en i onmen . Ab amo a
and o he s [2] di ide NoSQL da abases in o wo
ca ego ies: hose ha a e good in ead ope a ions and
hose ha a e good in upda e ope a ions. MongoDB,
Redis, and O ien DB belong o he i s ca ego y,
whe eas Cassand a and HBase belong o second
ca ego y. Cassand a again possessed he bes
pe o mance among column amily s o es and i
pe o med well especially in w i e ope a ions.
The e o e, Cassand a is a s ong al e na i e o inse
in ensi e applica ions.
Klein and o he s [17] ha e compa ed MongoDB
e sion 2.2, Cassand a e sion 2.0 and Riak e sion 1.4.
A modi ied e sion o he Yahoo! Cloud Se ing
Benchma k amewo k was used o es ing. Tes s
measu ed he h oughpu o ead-only, w i e-only and
ead/w i e wo kloads, and ead and w i e la encies. Each
es was un h ee imes wi h di e en numbe o
h eads. The pe o mance o Cassand a was clea ly bes
in he ead and w i e es s when he numbe o h eads
was inc eased. Riak had he second bes pe o mance
and hi d was MongoDB. On he one hand, Cassand a
had he bigges delay in ead and w i e ope a ions
whe eas Riak was 5 imes as e and MongoDB was 4
imes as e . The eason o he be e pe o mance o
Cassand a was ha i s hash based sha ding was much
mo e e icien han he sha ding o MongoDB. On he
o he hand, he indexing ea u es o Cassand a enabled
as que ies. Fu he mo e, he pee o pee based
a chi ec u e acili a ed e icien coo dina ion o ead and
w i e ope a ions be ween di e en nodes. F om he
pe spec i e o he p esen s udy, he esul s a e
in e es ing because we also un es s using he di e en
numbe o h eads.
All he men ioned es s, whe e Cassand a
pa icipa ed, we e key- alue o ien ed. Yahoo! Cloud
Se ice Benchma king ool was used in many o he
es s and his ool measu es ge and pu pe o mance
wi h di e en loads. As key- alue s o es a e well
op imized o hese kinds o que ies, hey had he bes
pe o mance. Howe e , among column amily s o es
Cassand a pe o med bes and he e o e we selec ed
Cassand a o es ing mul icolumn MNE da a. Table 1
summa izes he esul s o he pe o mance es s. Ou es
se ing di e s om abo e-men ioned e alua ion in wo
ways. Fi s , we use eal wo ds mul i-columned da a and
second, we e alua e no only NoSQL da abases bu also
compa e NoSQL wi h NewSQL da abases.
3 REVIEWING DATABASES FOR TESTING
In many o he cu en MNE applica ions, he solu ions
a e based on adi ional SQL da abases. As MNE da a
has a hea y demand o inse ion pe o mance, we a e
in e es ed in NoSQL and NewSQL solu ions o MNE
da a. In his pa , we e iewed di e en NoSQL da a
models and aimed o ind sui able da abases in o de o
in es iga e i adi ional SQL da abase could be eplaced
wi h he NoSQL and NewSQL solu ions o he needs o
MNE da a applica ions.
NoSQL da abases a e in ended o big da a se s and
hei o ganiza ion is no based on he ela ional model.
The que y capabili ies o NoSQL solu ions a e di e en
in compa ison wi h adi ional SQL solu ions, and his
mus be aken in o accoun when selec ing a sui able
NoSQL da abase solu ion o MNE da a ha is mul i-
column, s uc u ally p ede e mined and con ains a low
numbe o ela ionships. The g aph-based, column
amily s o es, documen s o es and key- alue s o es a e
di e en ypes o NoSQL da abases [12].
Key- alue s o e model is e y simple and does no
suppo as di e se que ies as SQL does. In mul icolumn
da a, SQL based solu ions a e di icul o eplace wi h
simple key- alue s o e solu ions because he keys o
que y mus be known be o ehand. G aph da abases a e
nei he a sui able choice o MNE da a because hey a e
designed o da a ha has a g ea numbe o
ela ionships. Ins ead, documen s o es and column
amily s o es a e sui able o MNE da a because bo h o
P. Ko i an a, M. Junkka i: E ec i eness o NoSQL and NewSQL Da abases in Mobile Ne wo k E en Da a: Cassand a and Pa S eam/Kine ic
5
Figu e 1: Radio Access Bea e in UMTS sys em
hem con ain s uc u ed da a unde key and suppo
que ies o di e en a ibu es o da a. Documen s o es
a e based on JSON o ma . Howe e , one o he
d awbacks o JSON is ha i consumes space because o
he s uc u e de ini ion o he JSON s anda d. Column
amily s o es, like Cassand a and Hype able, suppo
SQL-like que y language CQL (Cassand a Que y
Language) [6] and HQL (Hibe na e Que y Language)
[16]. Howe e , he exp ession powe o hese languages
is limi ed in compa ison wi h he s anda d SQL. Fo
example, join ope a ions a e no suppo ed.
In column amilies, da a also ake less space
compa ed wi h documen s o es. This is due he ac ha
column de ini ions ake less space on disc han JSON
s uc u e de ini ions. Thus, among NoSQL da abases,
he column amily s o e model sui s bes o he needs
o MNE da a. We es ed space consump ion o
1,000,000 RAB epo s in one o he mos popula
documen s o es, MongoDB and one o he mos popula
column amily s o es, Cassand a. The size o RAB
epo s as a MongoDB collec ion was 4.53 GiB and as a
Cassand a column amily was 1.18 GiB. As he sys em
migh ha e o s o e billons o hese epo s, we
conside ed Cassand a mo e space e icien .
Among column amily s o es, Cassand a has clea ly
he highes anking in DB-Engines [12]. In Cassand a
p e iously men ioned amoun o epo s consume
a ound as much space as in a SQL da abase. As
e iewed in Sec ion 2, Cassand a has he bes
pe o mance among he column amily s o es. The way
Cassand a s o es he da a should sui inse ions well.
When inse ing da a, Cassand a jus appends he da a
in o commi log and mem able. The ope a ion is simple
and hus inse ion ope a ions should be e icien .
The e o e, we selec Cassand a as a ep esen a i e o
column amily s o es in ou es ing.
NewSQL solu ions a e a new g oup o da abases ha
aim o p o ide he bes sides o he wo kinds o
da abases: he high scalabili y o NoSQL and he ACID
ea u es o adi ional ela ional da abases. The e a e
h ee main ca ego ies o NewSQL da abases: 1. New
da abase solu ions ha ha e been w i en om sc a ch;
2. MySQL based s o age engines; 3. Pluggable solu ions
o exis ing da abases ha aim o p o ide mo e
scalabili y [26]. We chose Pa S eam o ep esen a
NewSQL da abase wi h p ope ies om bo h SQL and
NoSQL da abases. Apa om suppo ing adi ional
SQL que ies, Pa S eam also p o ides ho izon al
scalabili y ha is no o e ed by adi ional SQL
solu ions. The Pa S eam da abase handles da a in
pa i ions, i.e. a able can be pa i ioned based on chosen
pa i ioning columns. This way Pa S eam enables e y
as que ying as i can exclude i ele an pa i ions by
using bi map indexing [24].
Fu he mo e, Pa S eam is one o he newes
NewSQL like da abase wi h possibili y o ins all a geo-
dis ibu ed analy ics se e . This is an essen ial ea u e
o geo-dis ibu ed elecommunica ion a chi ec u e.
Al hough we do no in es iga e geo-dis ibu ed analy ics
in he p esen s udy, his was s ill one eason o
selec ing Pa S eam ou o o he NewSQL solu ions. To
ou bes knowledge, no p e ious esea ch on Pa S eam
exis s, so we aim o show in he p esen esea ch how
his kind o da abase pe o ms agains Cassand a.
4 TEST DATA – RAB REPORTS
In a pe o mance e alua ion, i is essen ial ha he es
da a co esponds o eal da a [3]. Ou es da a
s uc u ally and in con en co esponds o Radio Access
Bea e (RAB) epo s. Radio Access Bea e s a e used
when a use equipmen , o example a mobile elephone,
connec s o a mobile ne wo k. RAB gua an ees
bandwid h o di e en kinds o communica ion ha a

Open Jou nal o Da abases (OJDB), Volume 5, Issue 1, 2018
6
mobile equipmen does in he ne wo k. Di e en so s
o RABs a e used o di e en ypes o communica ion.
Fo example, con e sa ional speech RABs a e used o
no mal elephone calls. These RABs gua an ee 12.2
kbps bandwid h o speech. Web b owsing and email
sending ac i i ies use in e ac i e packe swi ched RABs
ha gua an ee 384 kbps downlink and 64 kbps uplink.
Many o he kinds o RABs also exis o di e en kinds
o connec ion.
Figu e 1 illus a es he componen s o RAB in
UMTS sys em. As can be seen om he pic u e, Radio
Access Bea e consis s o Radio Bea e and Iu Bea e .
Radio Bea e is c ea ed be ween a use equipmen and
Radio Ne wo k Con olle (RNC). Radio Ne wo k
Con olle is an elemen ha is esponsible o managing
esou ces be ween a adio ne wo k and a co e ne wo k.
Iu Bea e is c ea ed be ween RNC and Mobile Swi ching
Cen e (MSC) in a ci cui swi ched co e ne wo k and
Se ing GPRS Suppo Node (SGSN) in a packe
swi ched co e ne wo k. MSC ou es oice calls and SMS
messages o he ci cui swi ched ne wo k. SGSN wo ks
simila ly o packe swi ched da a.
Ne wo k ope a o s a e in e es ed in moni o ing he
ac i i ies ha occu in he ne wo k. This is why a
ne wo k elemen , such as RNC, sends epo s when
RAB is c ea ed. The RAB epo is a pa o commonly
accep ed 3GPP (3 d Gene a ion Pa ne ship P ojec )
speci ica ions [28]. Ne wo k e en moni o ing sys ems
a e used o analyze he epo s. All he c ea ed RAB
epo s a e usually collec ed in o a sys em da abase. In
Figu e 1, da a a e collec ed om he RNC, MSC and
SGSN elemen s in o a da abase. Du ing ush hou s,
when a lo o RABs a e es ablished, a huge numbe o
epo s migh be sen , so i is e y impo an o he
da abase o pe o m as enough o handle all hese
epo s.
The con en o he RAB epo s a ies in some ex en
and depends on he ne wo k elemen ha sends hem. A
RAB epo ypically con ains in o ma ion on he use
equipmen ha eques s RAB and echnical in o ma ion
ela ed o a base s a ion and connec ion. In ou case, a
RAB epo con ains in o ma ion abou use equipmen
such as In e na ional Mobile Subsc ibe Iden i y (IMSI),
In e na ional Mobile S a ion Iden i y (IMEI), Mobile
S a ion In e na ional Subsc ibe Di ec o y Numbe
(MSISDN) and an IP add ess. The e is also much
in o ma ion ela ed o base s a ions and connec ions.
This in o ma ion includes s a and s op base s a ions
and hei Cell ID (CID), UTRAN Cell ID (LCID),
Mobile Coun y Code (MCC), Mobile Ne wo k Code
(MNC), Special A ea Code (SAC) and Loca ion A ea
Code (LAC). Fu he , RAB epo s con ain in o ma ion
abou many o he kinds o connec ion and possible
ailu e. The RAB epo s ha we used in ou es con ain
96 columns.
5 EVALUATION SETUP
A Ja a p og am was implemen ed o gene a e he es
da a ha s uc u ally co espond o eal RAB epo s
used in he UTRAN ne wo k elemen s. Mo e conc e ely,
he p og am c ea es an a ay objec wi h 96 columns o
a RAB epo , which con ains 88 columns o 32-bi
in ege ypes, one imes amp and se en s ing ypes. The
in ege s a e gene a ed using he andom class o Ja a
wi h he anges o eal alues wha a e used in he ac ual
epo s. Some s ing ypes, such as IMSI (In e na ional
Mobile Subsc ibe Iden i y) and URL (Uni o m
Recou se Loca o ), a e selec ed om he eal da a and
o he da a (e.g. IP 4 -In e ne P o ocol e sion 4) a e
gene a ed a andom. The imes amp is he ime when
he a ay is inse ed in o he da abase. All da a a e s o ed
in o a able called ne wo kda a. In Appendix A, he code
o gene a ing he es da a is gi en. Th ee columns we e
indexed ha co espond o eal indexing needs o
ypical use cases o MNE da a.
The compu e s used in he es s we e HP P oLian
DL380 Gen9 Se e . The used ope a ing sys em is Red
Ha En e p ise Linux e sion 6.5. Da as ax Cassand a
e sion 2.2 and Pa S eam e sion 3.3.4 we e ins alled.
Cassand a d i e e sion 2.1.5 was used o inse da a
in o Cassand a and Ja a S eaming Impo API e sion
3.3.4 was used o inse da a in o Pa S eam. We used
only one node ins alla ion o he bo h da abases. The HP
P oLian DL380 Gen9 Se e had ollowing ha dwa e
se ing:
 2 x In el Xeon E5-2667 3 CPU @ 3.20GHz
 8 co es, 16 h eads
 64-bi memo y echnology
 L1 cache 512 KB
 L2 cache 2048 KB
 L3 cache 20480 KB
 32 GB memo y o each p ocesso @ 2133 MHz.
In he es se ing, he se e s con ain only necessa y
p og ams and no unnecessa y ex e nal load exis ed. Fo
inse ing da a in o da abases, he Ja a p og am u ilizes a
o -loop ha i e a es h ough a lis o a ay objec s. The
alues o objec s a e andomly gene a ed ollowing he
da abase s uc u e p esen ed in Appendix A. In ou es s,
he ini ial size o da a was 10,000 ows. The o loops
ei he keep looping ough he lis un il ce ain amoun
o ime has passed o ce ain amoun o ows a e
inse ed. So as he same p e-gene a ed lis o ows is
looped ough many imes, some o he alues can be
duplica es. In ou es s, his is no impo an as we only
a e in e es ed in he inse ion and que y pe o mance.
The Ja a p og am can also inse alues in mul iple
h eads and he amoun o h eads can be selec ed.
The es s we e un in a single node o achie ing
compa abili y wi h he exis ing a chi ec u e ha is
designed o a single node da abase.
P. Ko i an a, M. Junkka i: E ec i eness o NoSQL and NewSQL Da abases in Mobile Ne wo k E en Da a: Cassand a and Pa S eam/Kine ic
7
Table 2: A e age inse ion a e in ows pe second
o wel e i e-minu e uns
Numbe
o Th eads
Pa S eam
Cassand a
1
30459
1132
2
56393
5454
3
79114
4918
4
101108
4491
5
117278
4032
6
130448
3619
7
151606
3161
8
167868
2885
9
197233
2492
10
220393
2187
11
242622
2063
12
241366
2029
6 EVALUATION
In his s udy, we e alua e he pe o mance o a NoSQL
da abase (Cassand a) and a NewSQL da abase
(Pa S eam), and aim a answe ing he ques ion whe he
new da abase solu ions b ing addi ional alue compa ed
o exis ing legacy SQL based solu ions. Since
applica ions o MNE da a a e inse in ensi e, his
e alua ion ocuses on he e iciency o s o ing and
agg ega ion que y p ocessing.
6.1 S o ing Speed
The e iciency o s o ing da a is essen ial in
elecommunica ion because du ing a sho ime pe iod a
la ge amoun o da a may be inse ed in o a da abase.
The numbe o h eads is a ypical way o inc ease
s o ing speed.
We i s compa e Pa S eam and Cassand a using a
single h ead. The a e age s o ing speed in Cassand a
was abou 12,500 ows pe second and in Pa S eam
abou 40,000 ows pe second. In o he wo ds, he
s o ing speed was o e h ee imes as e in Pa S eam
han in Cassand a. Du ing he wo-hou pe iod o es ing
he di e ence s ayed linea ly he same.
In o de o es he impac o he numbe o h eads
in s o ing da a, we made wel e i e-minu e uns using
di e en numbe o h eads wi h bo h da abases. The
esul s a e gi en in Table 2 and illus a ed in Figu e 2.
In Pa S eam, inc easing he numbe o h eads also
inc eased he s o ing speed. Cassand a, ins ead, did no
bene i no ably om inc easing he numbe o h eads.
The di e ence o he maximum s o ing speeds was 44
imes bigge in Pa S eam han in Cassand a. Based on
Figu e 2. A e age inse ion a e in ows pe second
wi h di e en numbe o h eads
he pe o mance o ela ional da abase solu ions, we
choose he inse ion speed o 210,000 ows pe second
as he pe o mance baseline. Cassand a did no exceed
he pe o mance baseline e en wi h he maximal numbe
o h eads, whe eas Pa S eam exceeded he
pe o mance baseline when using en o mo e h eads.
6.2 Que ying Speed
In MNE applica ions, i is o en needed o ind he
amoun o ce ain epo s. Thus, we chose a coun que y
o es ing he agg ega ion e iciency o da abases and
he coun que y is cu en ly used by exis ing SQL based
applica ions. The coun unc ion is hea y o da abases
o p ocess. I is suppo ed by bo h Pa S eam and
Cassand a and we we e hus able o compa e hem and
see he di e ences in pe o mance. In Cassand a, he use
o a column coun e is ecommended o keep ack o he
amoun o eco ds. Coun e columns wi h mul iple
nodes may in ol e consis ency issues.
The i s que y is a basic agg ega ion que y whe e
he numbe o all he ows is calcula ed.
Que y 1: SELECT COUNT(∗) FROM 𝑛𝑒𝑡𝑤𝑜𝑟𝑘𝑑𝑎𝑡𝑎
We es ed he speed o he que y o e di e en
amoun o da a. The es ed da a se s consis o hal
million, one million, en million and hund ed million
ows. The esul s a e gi en in Table 3. In all he cases,
Pa S eam was d ama ically as e han Cassand a. Fo
example, in case o he hal million ows, Cassand a
pe o med he que y in 36 seconds whe eas Pa S eam
used only ou milliseconds. In he case o hund ed
million ows, Cassand a could no inish he execu ion
o he que y du ing 30 minu es whe eas Pa S eam used
only 16 seconds.
I is also wo h no ing ha unning he same que y
wice imp o ed he execu ion ime in Pa S eam. In he
0
50000
100000
150000
200000
250000
300000
12345678910 11 12
inse ed ows
Numbe o h eads
Pa s eam Cassand a
Open Jou nal o Da abases (OJDB), Volume 5, Issue 1, 2018
8
Table 3: Execu ion ime o Que y 1
Numbe o Rows
Cassand a
Pa S eam
500,000
36 seconds
0.004 seconds (1s un)
0.002 seconds (2nd un)
1,000,000
1 min 24 seconds
0.007 seconds (1s un)
0.0035 seconds (2nd un)
10,000,000
11 min 12 seconds
0.067 seconds (1s un)
0.031 seconds (2nd un)
100,000,000
Timeou (30 minu es)
16 seconds (1s un)
0.35 seconds (2nd un)
cases o hal , one and 10 million ows, he ime hal ed.
In he case o hund ed million ows, he p ocessing ime
dec eased om 16 seconds o less han hal seconds.
This is due o he ac ha in he i s que ying he bi map
is loaded in o he cen al memo y, and hus i is
immedia e in use in he second que y p ocessing.
Cassand a does no bene i om epe i i ely que ying.
The abo e esul s do no mean ha Cassand a is
ine icien in gene al. I a que y is ocused on he
column ha is pa o he p ima y key, Cassand a is
e icien . Que y 2 ep esen s he que y ype whe e an
a ibu e is exac ly alua ed. We conside he que ying
e iciency ela ed o he ole o a alua ed a ibu e. The
es da a con ain a hund ed million ows.
Que y 2: SELECT COUNT(∗) FROM 𝑛𝑒𝑡𝑤𝑜𝑟𝑘𝑑𝑎𝑡𝑎
WHERE 𝑎𝑡𝑡𝑟𝑖𝑏𝑢𝑡𝑒 = 44755149
When he a ibu e belongs o a column ha is a pa
o he p ima y key in Cassand a, and he a ibu e is
pa i ioned and bi map indexed in Pa S eam, he e is no
signi ican di e ence be ween he que y pe o mances
o he wo da abases. Execu ion imes o wo da abases
we e less han 0.01 seconds. I he a ibu e is indexed
bu no pa i ioned in Pa S eam, hen he que y was
pe o med in 18 seconds. I he a ibu e is nei he
indexed no pa i ioned in Pa S eam, he que y equi ed
mo e han 40 seconds. We canno un he que y wi h
hese se ings in Cassand a, because non-indexed
a ibu es canno be que ied by Cassand a.
The las que y is s ing ma ching, which is suppo ed
by Pa S eam bu no by Cassand a. In Que y 3, ‘515’ is
an a ea code ha is he ini ial code o IMSI ( he
In e na ional Mobile Subsc ibe Iden i y).
Que y 3: SELECT COUNT(∗) FROM 𝑛𝑒𝑡𝑤𝑜𝑟𝑘𝑑𝑎𝑡𝑎
WHERE imsi LIKE ′515%′
In es ing Que y 3, he da abase con ained en
million ows. Pa S eam pe o med he s ing ma ching
que y in 10 seconds when he bi map was no used,
whe eas when using he index, he que y equi ed 28
seconds. This means ha bi map indexing does no
inc ease pe o mance o e e y ype o que ies.
7 DISCUSSION
We e alua ed he inse ion pe o mance and he
agg ega ion capabili y o Pa S eam and Cassand a
da abases o e MNE da a in he p e ious sec ion. The
esul s o he e alua ion show ha Pa S eam was much
mo e e icien in inse ion speed compa ed wi h
Cassand a.
In o de o ind ou why Cassand a had an in e io
inse ion speed, we pe o med mo e in es iga ions on i s
pe o mance by changing he es se ing. We changed
du able w i es ea u e o o -s a e ha bypasses
commi log o Cassand a. Cassand a appends he da a
i s in o commi log- ile and hen akes i in o in-
memo y mem able. I mem able is ull, da a a e s o ed
in o da a iles. Howe e , se ing du able w i es o o -
s a e did no ha e any impac on inse ion speed. We
also es ed he e ec o di e en sizes o he mem able,
bu his did no imp o e inse ion pe o mance ei he .
Pa S eam seems o be be e op imized o pa allel
inse ions han Cassand a. The e a e di e ences o how
he da a is s o ed and he da abases a e implemen ed.
Cassand a is p og ammed in Ja a and canno be
op imized as e icien ly as C/C++ p og am. Cassand a
s o es da a as key- alue pai s. All he alues a e s o ed
wi h hei co esponding key. When mo e columns we e
indexed, inse ion speed seemed dec easing. Thus,
w i ing da a as key- alue pai s seems o be one o he
ac o s ha dec ease pe o mance.
In he agg ega ion es , Cassand a was clea ly
in e io o Pa S eam as well. When execu ing he coun
unc ion o CQL (Cassand a Que y Language) [6],
Cassand a eads h ough all he ows in he da abase and
he ope a ion is e y slow. I an agg ega ion que y is
modi ied such ha a condi ion o an indexed key is
inse ed, bo h o Cassand a and Pa S eam show equally
good pe o mance when he que ied alue belongs o an
indexed column. This was o be expec ed. Cassand a
e ches he columns quickly wi h a igh ow key and
Pa S eam akes ad an age o pa i ioning and bi map
indexing. I he column is no pa i ioned, Pa S eam
canno exclude i ele an pa i ions om he que y. This
P. Ko i an a, M. Junkka i: E ec i eness o NoSQL and NewSQL Da abases in Mobile Ne wo k E en Da a: Cassand a and Pa S eam/Kine ic
9
is why he que y akes a longe ime o inish. When
using bi maps he que ying ime will be hal ed.
Que y 3 was execu ed jus o Pa S eam as
Cassand a does no suppo bi map indexing and CQL
does no suppo LIKE-ope a ion. The que y
demons a es ha bi map indexing does no always
p o ide be e pe o mance. This que y was as e when
bi map indexing was no used. Seeking he alues ha
ma ch a like pa e n seems o be he weakness o he
bi map indexing in Pa S eam. Que y 3 exp esses also a
gene al p oblem in compa ing di e en da abases by
complex que ies. Namely, i a da abase o a que y
language does no suppo a que y ype, he compa ison
canno be execu ed. This is one eason o de eloping
SQL++ [22] ha gi es a simila in e ace o ela ional
da abases and NoSQL da abases. On he o he hand, he
de elopmen o he que y languages o NoSQL
da abases is in p og ess and, hus, he SQL++ in e aces
will be de eloped.
Al hough Pa S eam seems o be o e whelming in
da a inse ion and agg ega ion o e NME da a, an open
ques ion is he e iciency o Pa S eam in gene al. Fo
ha , Pa S eam should be es ed in di e en da a se s
and compa ed wi h o he NoSQL and NewSQL
da abases. Yahoo! Cloud Se ice Benchma king ool
would ge esul s ha will be in line wi h he esul s
eleased in NoSQL da abases. TCP-H and TCP-BB [5]
a e public da a en i onmen s wi h s uc u ally mo e
complex da a. The Pa S eam has been acqui ed by
Cisco and in eg a ed in o Cisco Kine ic dis ibu e
sys em in he summe 2018, and hus u he es s should
be ocused on Kine ic. Du ing ou es pe iods we had
only he license o Pa S eam bu no Kine ic. In gene al,
a simila es se ing can be epea ed wi h Kine ic, and
simila esul s could be expec ed wi h Kine ic.
8 CONCLUSIONS
In his wo k, we in es iga e he applicabili y o NoSQL
and NewSQL da abases o s o ing and que ying Mobile
Ne wo k E en (MNE) da a. S uc u ally, column
amily s o es and documen s o es a e sui able o MNE
da a, bu he s o age o ma used by documen s o es
consumes space. The e o e, among NoSQL da abases
we selec ed he column amily s o e da abases o
pe o mance e alua ion. We es ed he pe o mance o
wo da abases o e MNE da a: Cassand a and
Pa S eam. Cassand a is a column amily s o e da abase
and i is known as a e y e icien solu ion o big da a
se s. Pa S eam is a NewSQL like da abase o which no
es esul s ha e so a been published.
MNE applica ions a e inse in ensi e and he e o e
he e iciency o s o ing da a is essen ial in MNE
applica ions. In e ms o a single h ead, Pa S eam was
h ee imes as e han Cassand a. When inc easing he
numbe o h eads, bo h da abases enhanced hei
s o ing pe o mance, bu Pa S eam inc eased ob iously
mo e. The di e ence o he maximum s o ing speed was
44 imes bigge in Pa S eam han in Cassand a. In he
es se ing ee columns we e indexed. This seems o be
he essen ial eason o he huge di e ence in he
pe o mance. In an addi ional es , we ound ha i only
one column is indexed, Pa S eam was six imes as e
han Cassand a.
In agg ega ion que ying, Pa S eam was
d ama ically as e han Cassand a. When a que y is
ocused on a key a ibu e, no di e ence be ween he
da abases was ound. So a Cassand a is known as one
o he column amily s o es wi h bes pe o mance
especially o w i e ope a ions. Howe e , wi h inse -
in ensi e MNE applica ions, Pa S eam is e y e icien
compa ed o Cassand a. The unc ionali y o Pa S eam
is now a pa o Cisco Kine ic and simila esul s could
be expec ed wi h Cisco Kine ic
The s udy p esen ed in his pape clea ly indica es
ha he new kinds o da abase echnology can clea ly
b ing pe o mance ad an ages o e legacy da abase
echnology and o e a e y s ong al e na i e o exis ing
solu ions.
REFERENCES
[1] Y. Abubaka , T.S. Adeyi and I.G. Au a,
“Pe o mance e alua ion o NoSQL sys ems
using YCSB in a esou ce aus e e en i onmen ,”
In . J. o Appl. In . Sys ems, ol 7, pp. 23-27, 2014.
[2] V. Ab amo a, J. Be na dino, and P. Fu ado,
“Which NoSQL da abase? A pe o mance
o e iew,” Open J. o Da abases (OJDB), ol. 1,
no. 2, pp. 17-24, 2014. [Online]: h p://nbn-
esol ing.de/u n:nbn:de:101:1-201705194607
[3] A. Alexand o , C. B ücke and V. Ma kl, “Issues
in big da a es ing and benchma king,”
in P oceedings o Six h In e na ional Wo kshop
on Tes ing Da abase Sys ems, 2013.
[4] S. Bushik, “A endo -independen compa ison o
NoSQL da abases: Cassand a, HBase, MongoDB,
Riak,” Ne wo k Wo ld, 2012.
[5] P. Cao, B. Gowda, S. Lakshmi, C.
Na asimhade a a, P. Nguyen, J. Poelman, "F om
BigBench o TPCx-BB: S anda diza ion o a big
da a benchma k," In P oceedings o Technology
Con e ence on Pe o mance E alua ion and
Benchma king, pp. 24-44, 2016.
[6] Cassand a, “The Cassand a Que y Language
(CQL),” Apache So wa e Founda ion, online:
h p://cassand a.apache.o g/doc/4.0/cql/,
accessed Augus 20, 2018.