senso s
A icle
Vi ualized MME Design o IoT Suppo in
5G Sys ems
Pila And es-Maldonado *,†, Pablo Ameigei as †, Jona han P ados-Ga zon †,
Juan Jose Ramos-Munoz †and Juan Manuel Lopez-Sole †
Depa men o Signal Theo y, Telema ics, and Communica ions, Uni e si y o G anada, G anada 18071, Spain;
pameigei as@ug .es (P.A.); jpg@ug .es (J.P.-G.); jj amos@ug .es (J.J.R.-M.); juanma@ug .es (J.M.L.-S.)
*Co espondence: pam91@co eo.ug .es; Tel.: +34-647-328-885
† These au ho s con ibu ed equally o his wo k.
Academic Edi o s: Ing id Moe man, Je oen Hoebeke and Eli De Poo e
Recei ed: 22 May 2016; Accep ed: 16 Augus 2016; Published: 22 Augus 2016
Abs ac :
Cellula sys ems a e ecen ly being conside ed an op ion o p o ide suppo o he In e ne
o Things (IoT). To enable his suppo , he 3 d Gene a ion Pa ne ship P ojec (3GPP) has in oduced
new p ocedu es speci ically a ge ed o cellula IoT. Wi h one o hese p ocedu es, he ansmissions
o small and in equen da a packe s om/ o he de ices a e encapsula ed in signaling messages and
sen h ough he con ol plane. Howe e , hese ansmissions om/ o a massi e numbe o de ices
may imply a majo inc ease o he p ocessing load on he con ol plane en i ies o he ne wo k and
in pa icula on he Mobili y Managemen En i y (MME). In his pape , we p opose wo designs o
an MME based on Ne wo k Func ion Vi ualiza ion (NFV) ha aim a acili a ing he IoT suppo .
The i s p oposed design pa ially sepa a es he p ocessing esou ces dedica ed o each a ic class.
The second design includes a ic shaping o con ol he a ic o each class. We conside h ee
classes: Mobile B oadband (MBB), low la ency Machine o Machine communica ions (lM2M) and
delay- ole an M2M communica ions. Ou p oposals enable educing he p ocessing esou ces and,
he e o e, he cos . Addi ionally, esul s show ha he p oposed designs lessen he impac be ween
classes, so hey ease he compliance o he delay equi emen s o MBB and lM2M communica ions.
Keywo ds: NFV; i ualiza ion; 5G; LTE; M2M; IoT; a ic peaks
1. In oduc ion
The In e ne o Things (IoT) is a e m used o a se o echnologies, sys ems and de ices ha
enable connec i i y o he In e ne and ha a e based on he physical en i onmen [
1
]. Wi h a wide
ange o po en ial applica ions, IoT de ices a e apidly sp eading, and o ecas s a e p edic ing a huge
g ow h o hese de ices o e he nex ew yea s [2].
Recen ly, cellula sys ems a e being conside ed as an op ion o p o ide connec i i y o IoT
de ices due o hei ubiqui ous p esence, widesp ead co e age, eliabili y and suppo o mobili y [
3
].
Wi hin he cellula con ex , he IoT connec i i y solu ion is e e ed o as Machine- o-Machine (M2M).
Howe e , cellula sys ems ha e been designed o Human o Human communica ions (H2H) o
Mobile B oadband Access (MBB), and consequen ly, he widesp ead p o ision o M2M se ices wi h
hese sys ems en ails signi ican echnical challenges [
4
]. Some o hese challenges a e: he scalabili y
issues aised by he huge numbe o expec ed de ices; he ime- a ying a ic cha ac e is ics o M2M
applica ions, which a e e y di e en om MBB o H2H; and he e y he e ogeneous QoS demands in
e ms o bandwid h, la ency and eliabili y.
The 3 d Gene a ion Pa ne ship P ojec (3GPP) has included enhancemen s in Long-Te m
E olu ion (LTE) ne wo ks o he deploymen o IoT. One o hem is da a anspo in Con ol
Plane Cellula IoT E ol ed packe sys em Op imiza ion (CPCEO) [
5
]. This is a se o ansmission
Senso s 2016,16, 1338; doi:10.3390/s16081338 www.mdpi.com/jou nal/senso s
Senso s 2016,16, 1338 2 o 25
p ocedu es speci ically designed o small and in equen da a ansmissions om/ o M2M de ices.
CPCEO p ocedu es employ Non-Access S a um (NAS) messages o ans e da a packe s om/ o he
de ice, ins ead o he es ablishmen and u iliza ion o da a plane bea e s.
The adop ion o he CPCEO p ocedu es can mi iga e ele an issues caused by small and
in equen da a packe ansmissions by a huge numbe o M2M connec ed de ices. CPCEO p ocedu es
educe he signaling explosion gene a ed by he es ablishmen and elease o da a bea e s.
This educ ion a ec s he co e and especially he adio in e ace due o he limi ed adio esou ces.
Howe e , CPCEO p ocedu es imply a majo inc ease o he p ocessing load on he con ol plane o he
E ol ed Packe Co e (EPC) and, in pa icula , on he Mobili y Managemen En i y (MME). In addi ion,
he cu en exposi ion o he EPC en i ies o he signaling in LTE [
6
], combined wi h he ixed capaci y
o cu en co e LTE ha dwa e-based in as uc u e, can limi he scalabili y o he CPCEO solu ion.
Cu en ly, majo e o s a e being made o esea ch and de elop new echnologies o u u e
5G sys ems [
7
]. Ne wo k Func ion Vi ualiza ion (NFV) is one o he p omising new echnologies
o 5G. NFV p o ides a no el amewo k o deploy ne wo k se ices on o i ualized se e s. The
use o he NFV pa adigm o i ualize he EPC en i ies and, in pa icula , he MME could acili a e
he wide deploymen o M2M communica ions in 5G. I imp o es he scalabili y and lexibili y o
he ne wo k compa ed o ha dwa e-based en i ies. This bene i is c ucial o he o eseen signaling
explosion gene a ed by M2M connec ed de ices. Due o he p omising bene i s o NFV, he e a e wo ks
ha ha e ackled he a chi ec u al and implemen a ion issues o applying his i ualiza ion pa adigm
o he LTE EPC; see, o example, [
8
–
11
]. Howe e , hese p oposals do no speci ically conside suppo
o M2M a ic.
In his pape , we p opose wo designs o a i ualized MME ha speci ically aim a acili a ing
he IoT suppo in 5G sys ems. The i s p oposed design pa ially sepa a es he p ocessing esou ces
dedica ed o di e en a ic classes; while he second design includes a ic shaping o con ol he
a ic o each class. In ou sys em, we ha e conside ed MBB, low la ency M2M and delay- ole an
M2M a ic classes. Ou p oposed schemes adjus he p ocessing esou ces conside ing he a ic
classes o se e and hei QoS equi emen s.
We ha e analyzed he unning cos s o he esou ces needed and he delay pe o mance. We ha e
compa ed ou p oposed schemes o wo o he schemes: (i) a baseline i ualized MME design, which
does no apply any speci ic a ic ea men pe class; (ii) an o e dimensioned i ualized MME. Fo
he cos analysis, we ha e conside ed a heo e ical model o dimension he equi ed esou ces, and
he da a cen e se up and billing model o he Amazon Elas ic Compu e Cloud (EC2). Fo he delay
pe o mance e alua ion, we ha e simula ed each scheme.
The esul s show ha ou schemes p o ide much lowe cos s han he o e dimensioned one.
Fu he mo e, hey p o ide simila delay pe o mance o MBB and low la ency M2M communica ions.
The ob ained delay sa is ies he exigen equi emen s o MBB and low la ency M2M communica ions.
Howe e , ou p oposed schemes apply a speci ic delay equi emen pe a ic class. This allows he
sa ing o p ocessing esou ces o delay- ole an M2M a ic.
The emainde o his pape is o ganized as ollows. Sec ion 2gi es an in oduc ion o he main
opics on which his wo k is based and poses he add essed p oblem. Sec ion 3in oduces a de ailed
desc ip ion o he sys em model and assump ions made. The adop ed a ic models a e explained in
Sec ion 4. In Sec ion 5, we explain he conside ed schemes o s udy, including ou p oposed schemes.
Sec ion 6p esen s he queue model used in he dimensioning o he i ualized MME. Sec ion 7
analyzes he esul s. Finally, he conclusion is in Sec ion 8.
2. Backg ound
2.1. In e ne o Things
The con e gence o he digi al and he physical wo ld p o ided by he In e ne o Things (IoT)
allows he in e ac ion be ween he en i onmen and he de ices connec ed o he ne wo k ha collec
Senso s 2016,16, 1338 3 o 25
in o ma ion. Unde his umb ella, many applica ions can be en isaged, which ha e led o se e al
ma ke s ( e icals) o IoT. In IoT applica ions, machines (o de ices) connec ed o he ne wo k can
communica e among hem, o wi h he applica ion se e , wi hou human in e ac ion. This is known
as Machine- o-Machine (M2M) communica ions.
IoT applica ions a e designed o speci ic e icals, such as indus y (e.g., moni o ing indus ial
plan s, boa ding ope a ion), ene gy (e.g., in en o y, was e collec ion), au omo i e (e.g., bike sha ing,
pa king sys em), heal hca e (e.g., diagnos ics, mobile assis ance) o media and en e ainmen
(e.g., paymen sys ems, com o able li ing)
[
12
]. These e icals ha e di e en equi emen s ha
de e mine IoT applica ions’ design, as o example, he pa icula connec i i y needs, de ice ene gy
consump ion, capabili ies, secu i y o a ic cha ac e is ics.
Despi e he e being many possible equi emen s o IoT applica ions, M2M communica ions
used in IoT a e being classi ied by o ganiza ions, such as he 3GPP o he Mobile and wi eless
communica ions Enable s o he Twen y- wen y In o ma ion Socie y (METIS), in wo big ends [
13
],
based on hei Quali y o Se ice (QoS) equi emen s:
•
Massi e M2M (mM2M): cha ac e ized by low cos /ene gy de ices, small da a olumes and a
massi e numbe o de ices connec ed. Among o he s, ele an examples a e sma me e ing,
lee acking o building au oma ion.
•
Low la ency M2M (lM2M): cha ac e ized by s ic equi emen s, such as ul a eliabili y, low
la ency and high a ailabili y. The dis inc equi emen s depend on he speci ic applica ion.
Fo ins ance, he end- o-end la ency can each down o a ew milliseconds o e en lowe [
14
].
To achie e ha , low la ency se ices could be se ed by di e en co e en i ies in he ne wo k. Fo
simplici y, we assume ha he lM2M de ices conside ed in his wo k can be handled by he MME.
Indus ial moni o iza ion, secu i y o heal hca e, and ac ic In e ne a e examples o his end.
Fo mos o M2M communica ions used in IoT applica ions, small and in equen da a
ansmissions a e a common a ic cha ac e is ic, as men ioned in [
15
]. Ne e heless, his a ic
cha ac e is ic is handled ine icien ly by cu en mobile ne wo ks, such as LTE. This is mainly due
o he p io esou ce ese a ion equi ed be o e he ansmission [
16
]. To sol e ha , he 3GPP has
included LTE enhancemen s o M2M in i s speci ica ions, ecen ly in eg a ed in o he con ex o
Cellula IoT (CIoT). Fo example, he 3GPP has in oduced in [
5
] new p ocedu es o educe signaling o
ans e da a packe s om o o he de ice, named da a anspo in Con ol Plane CIoT E ol ed packe
sys em Op imiza ion (CPCEO). To his end, he p ocedu es use Non-Access S a um (NAS) anspo
capabili ies. In LTE, NAS is used o ans e non- adio signaling be ween he Use Equipmen (UE) and
he MME. In he ansmission o uplink da a, wi h his new p ocedu e, he da a packe is sen as NAS
signaling by he UE o he MME a e Radio Resou ce Con ol (RRC) es ablishmen . Nex , he MME
ca ies ou he secu i y check, and hen, i o wa ds he da a packe o he Se ing Ga eway (SGW)
(see Figu e 1). The CPCEO p ocedu es a oid he es ablishmen and elease o da a plane bea e s o
small and in equen packe s om M2M de ices, bu he MME has o p ocess hese packe s om he
da a plane ansmi ed as NAS signaling.
In he LTE E ol ed Packe Co e (EPC) a chi ec u e, he MME deals wi h he con ol plane.
Tha is, his en i y is he main signaling node in he EPC. I s main unc ions a e: NAS signaling,
use au hen ica ion, mobili y managemen (e.g., paging, use acking) and bea e managemen . Fo
mo e in o ma ion abou he MME en i y, see [
5
]. The e o e, he main d awback o hese new p ocedu es
o cellula IoT is he inc eased p ocessing capaci y equi ed by he MME [17].
In addi ion o his ans e op imiza ion, a new na owband adio echnology o IoT was
de eloped in Release 13, called he Na owband In e ne o Things (NB-IoT). The objec i e o NB-IoT
is o educe complexi y, inc ease co e age, achie e long ba e y li e ime and suppo a massi e numbe
o de ices. To achie e hese objec i es, NB-IoT has low da a a es, a 164-dB maximum coupling loss
a ge o s andalone mode ope a ion and limi ed mobili y suppo , and i allows only he hal -duplex
equency-di ision duplexing ope a ion. NB-IoT equi es a educed bandwid h o 180 kHz o bo h
Senso s 2016,16, 1338 4 o 25
downlink and uplink and u he p o ocol op imiza ions [
18
]. Fo UEs ha suppo he NB-IoT,
he 3GPP ag eed ha he con ol plane CIoT E ol ed Packe Sys em op imiza ion ea u e explained
abo e will be manda o y [5]. Fo mo e in o ma ion abou NB-IoT, see [19].
Figu e 1. Mobile-o igina ed da a anspo in he NAS packe da a uni [5].
2.2. Vi ualiza ion
NFV is one o he enabling echnologies o 5G sys ems, because i will allow ne wo k ope a o s o
cope wi h challenges, such as he a ic inc ease and IoT suppo while educing CAPEX and OPEX.
Wi h NFV, ne wo k unc ionali ies a e implemen ed wi h Comme cial O -The-Shel (COTS) ha dwa e,
which is much less expensi e han he e y specialized ha dwa e used in 4G ne wo ks. In his way,
ope a o s can c ea e, scale and deploy ne wo k componen s whene e hey a e needed, acco ding
o he pa icula eal- ime a ic condi ions. These bene i s a e ele an o he deploymen o IoT
in cellula ne wo ks. NFV could acili a e he adap a ion o he ne wo k en i ies o he new demand.
Wi h NFV, he in as uc u e se ices can be p og ammed, ins ead o e-a chi ec ing he in as uc u e
o he ne wo k [20].
In [
9
], he au ho s discuss he design o a i ualized EPC o implemen a ion in a cloud compu ing
en i onmen and i s o e ing “as a Se ice” (EPCaaS). They p esen and analyze a numbe o di e en
implemen a ion op ions o he i ualized EPC. Focusing on he 1:N mapping op ion, each en i y
o he EPC is decomposed in o h ee logical componen s: a F on End (FE), Se ice Logic (SL) and
S a e Da abase (SDB). The FE ac s as a communica ion in e ace wi h o he en i ies o he ne wo k, o
example he FE e mina es S eam Con ol T ansmission P o ocol (SCTP) sessions be ween eNodeB and
MME [
21
] and balances he load among se e al SLs. SLs implemen he p ocessing. The SDB s o es he
use session s a e, making he SLs s a eless. This design ollows he mul i- ie ed web se ices pa adigm
o cloud-based applica ions. Each o he h ee logical componen s is a ie o he mul i- ie a chi ec u e.
Senso s 2016,16, 1338 5 o 25
SLs a e s a eless, and he e o e, hey can scale ou /in wi hou impac ing ex e nally-connec ed pee s
and independen ly om o he SLs. The scaling o SDB and FE, which a e s a e ul componen s, is
possible [22], al hough i equi es a ca e ul design.
In mul i- ie ed web se ices, dedica ed hos ing e e s o he case whe e a se e is alloca ed o
a mos one applica ion a any gi en ime [
23
]. On he con a y, in sha ed hos ing, mul iple small
applica ions sha e each se e . Dedica ed hos ing is used o unning la ge clus e ed applica ions whe e
se e sha ing is un easible due o he wo kload demand imposed on each indi idual applica ion.
Dimensioning is he classic p ocess o de e mine he amoun o esou ces equi ed o p ocess
a se o eques s wi h a speci ic Quali y o Se ice (QoS) a ge . Howe e , dimensioning is a s a ic
p ocess, whe eas he a ic in mobile ne wo ks s ongly luc ua es. Howe e , dynamic p o isioning
o esou ces acco ding o he luc ua ions in he eques a e is easible hanks o he i ualiza ion
echnologies and he wide adop ion o cloud compu ing sys em in as uc u es [
24
]. As in he case
o In e ne wo kloads, mobile ne wo k a ic exhibi s long- e m a ia ions, such as ime-o -day o
seasonal e ec s, as well as sho - e m luc ua ions caused by unexpec ed e en s. The e o e, i is
easonable o conside applying dynamic p o isioning echniques o mul i- ie ed web se ices o
i ualized EPC en i ies.
In [
22
], he au ho s s udy he p o ision esou ces in mul i- ie applica ions and p opose a s a egy
o a oid bo leneck shi ing. They p opose o b eak down he end- o-end esponse ime
D
in o pe - ie
esponse imes D1,D2,... Dj, such ha :
j
∑
j=1
Dj=D. (1)
Then, hey p opose o de e mine how many se e s o alloca e o each ie sepa a ely, such ha
ie
j
p ocesses he eques s wi h a se ice esponse ime equal o
Dj
. Fo his capaci y p edic ion,
an es ima e o he peak session a e is used. In Sec ion 5, we will use his s a egy in he es ima ion o
esou ces o he i ualized MME en i y.
We will u he assume ha he i ualized EPC en i ies a e deployed using a cloud p o ide
se ice, such as Amazon’s Elas ic Compu ing Cloud (EC2), in which he compu ing esou ces can
be dynamically scaled in o ou on demand and wi h a pay-as-you-go p ice scheme. Fo example,
Amazon EC2 cha ges a pe -hou p ice o he se ice use [
25
]. Addi ionally, a ele an aspec in cloud
se ices such as EC2 is ha he ime equi ed o ob ain and boo a new se e ins ance can be in he
ange o ens o seconds o minu es [26].
2.3. Cha ac e is ics o M2M T a ic in Cellula Ne wo ks
The pene a ion o connec ed machines in he nea u u e is expec ed o ou numbe human
connec ions, eaching alues up o one o de o magni ude abo e [
27
]. This numbe and he a ic
cha ac e is ics o M2M de ices make di icul hei wide deploymen in cu en cellula ne wo ks, such
as LTE [28].
M2M communica ions ha e di e en cha ac e is ics compa ed o con en ional human-based
communica ions. M2M a e age a ic olume is small, abou wo o de s o magni ude lowe han
MBB a ic olume. The M2M a io o uplink a ic olume is highe han downlink a ic olume.
Fu he mo e, synch onized communica ion om a la ge numbe o de ices causes a ic peaks,
e.g., in pe iods o 1 h, 30 min o 15 min, as shown in he ime se ies analysis o M2M a ic [
29
].
Besides epo ansmission synch oniza ion e en s, he e will be ala ms caused by unp edic able
e en s. Toge he , all o hese e en s could imply peaks o signaling load wo- o ou - imes highe han
he a e age load [30]; see Figu e 2.
F om he ne wo k poin o iew, he scalabili y o he con ol p ocedu es igge ed by hese
ansmissions is one o he mos impo an challenges o M2M suppo . As wi h he CPCEO, he MME
is o con ey hese da a packe s; i has o be designed o be able o deal wi h his a ic inc ease and o
a oid any associa ed conges ion. Dynamic p o isioning o p ocessing esou ces is o li le use o he

Senso s 2016,16, 1338 6 o 25
MME o ca y he a ic caused by hese a ic peaks. Fi s , he billing models o cloud p o ide s
(such as Amazon EC2) ypically cha ge pe hou , and a ic peaks usually occu a sho e ime
in e als. Second, he e will be M2M a ic peaks caused by ala m e en s ha canno be p edic ed,
and he e o e, hey would cause conges ion du ing he ens o seconds o minu es needed o boo new
se e ins ances. Fo hese easons, we will assume ha he amoun o esou ces equi ed o se e
hese a ic peaks has o be dimensioned in ad ance.
Time
T a ic
Figu e 2. Example o M2M a ic peaks.
3. Sys em Model
In his wo k, we assume an access cellula ne wo k a chi ec u e based on LTE, which p o ides
se ice o Use s Equipmen (UE) and Machine-Type Communica ions De ices (MTCDs). Al hough
his a chi ec u e assumes he en i ies de ined in LTE/EPC, i could also be ex ended o o he cellula
a chi ec u es. The o e all sys em model is depic ed in Figu e 3.
E-UTRAN
eNB
PDN
eNB
SGW PGW
UE
MTCD
UE
MTCD
FE SDB
SL1
SL2
SL3
F on end
(Load balance )
S a eless MME
SL ins ances
S a e da abase
MME
Signaling
Da a
Figu e 3. O e all sys em model.
The main en i ies a e explained nex . Addi ionally, he no a ion and main de ini ions used in his
wo k a e summa ized in Table 1.
Senso s 2016,16, 1338 7 o 25
Table 1.
P ima y de ini ions. FE, F on End; SL, Se ice Logic; SDB, S a e Da abase; MBB, Mobile
B oadband; UE, Use Equipmen ; mM2M, massi e M2M; MO-CPCEO, Mobile-O igina ed Con ol
Plane Cellula IoT E ol ed packe sys em Op imiza ion.
No a ion Desc ip ion
TjMean esponse ime o each ie (whe e j∈{FE,SL,SDB})
DjTa ge mean esponse ime o ie j
NMBB Numbe o MBB UEs
NhNumbe o each class o M2M de ices (whe e h∈{mM2M,lM2M})
λSR Mean gene a ion a e pe MBB UE o Se ice Reques
λSRR Mean gene a ion a e pe MBB UE o Se ice Release
λHO Mean gene a ion a e pe MBB UE o Hando e
λMBB Mean a i al a e o con ol messages gene a ed by MBB UEs
λac i e Mean M2M de ice packe a e in he ac i e s a e
λala m Mean M2M de ice packe a e in he ala m s a e
ahPe cen o M2M de ices o each class ha changes o he ala m s a e in an M2M e en
phPe cen o each class o M2M de ices
pe
hPe cen in a ce ain ins an o ime an M2M de ice is in he ala m s a e
Ra io o M2M de ices pe MBB UE
λ
h
Mean a i al a e o MO-CPCEO pe M2M class
in a obse a ion pe iod (whe e ∈{ac i e,e en ,maxe en })
λhMean a i al a e o MO-CPCEO con ol messages gene a ed pe M2M class
whWeigh ing ac o o a i al a es o he absence o no o M2M e en s pe M2M class
λ
hTa ge a i al a e o dimensioning
λin smoo h
hTa ge a i al a e wi h in ense smoo hing o peaks
λsmoo hpeak
hTa ge a i al a e wi h mode a e smoo hing o peaks
λpeak
hTa ge a i al a e o peak a ic
λFE F on end a ge a i al a e
λSL Se ice logic a ge a i al a e
λSDB S a e da abase a ge a i al a e
µjSe ice a e o ie ins ance j
mjNumbe o ins ances o ie j
3.1. The Use Equipmen
UEs a e he e minals ha allow each human use o connec o he ne wo k ia he eNodeB base
s a ions. The UEs un he use s’ applica ions, which gene a e o consume ne wo k a ic (see he
a ic model in Sec ion 4). I a UE is in he idle s a e and i has da a o ansmi , i ca ies ou a Se ice
Reques (SR) p ocedu e o es ablish a bea e , which will con ey he da a packe s. We assume ha
UEs mo e ollowing a luid- low mobili y model. When a UE c osses he bo de be ween wo cells,
i igge s an X2-based Hando e (HO) p ocedu e.
3.2. MTC De ices
We assume ha MTCDs a e placed in ixed loca ions, and hey send small da a packe s ( epo s)
in equen ly o cen alized se e s. Following he 3GPP and METIS guidelines (see Sec ion 2.1),
we conside wo ypes o MTCDs, mM2M and lM2M de ices. We assume mM2M de ices un
delay- ole an M2M applica ions, which can ole a e seconds o delay [
31
], such as sma me e ing
o ag icul u e applica ions. We also assume ha lM2M de ices un s ic M2M applica ions, such as
indus ial applica ions o heal hca e, which a e cha ac e ized by low la ency equi emen s among
o he demands. We also assume a la ge numbe o MTCDs connec ed o he ne wo k, which will
Senso s 2016,16, 1338 8 o 25
cause a ic peaks due o M2M synch oniza ion o ala m e en s. Fo simplici y, we assume he e is no
coo dina ion among mM2M and lM2M e en s.
As he M2M a io o uplink a ic olume is highe han downlink (see [
29
]), we only conside
uplink M2M a ic. To ansmi a epo , we assume ha an M2M de ice igge s he Mobile-O igina ed
da a anspo signaling p ocedu e, de ined in he Con ol Plane CIoT E ol ed packe sys em
Op imiza ion (MO-CPCEO) [
5
]. Rega ding he MO-CPCEO p ocedu e, we assume ha he ollowing
con ol messages do no ake place (see Figu e 1):
•
Messages 4 o 7 a e no used, since he e is a connec ion es ablished be ween MME and SGW,
and he e is no impo an change o in o m he Packe da a ne wo k Ga eway (PGW);
•Messages 9 o 13 a e no used as, no downlink da a a e expec ed by he MTCD.
Then, he MME has o p ocess one con ol message o each MO-CPCEO p ocedu e.
3.3. eNodeB S a ions
eNodeB (eNB) s a ions ecei e signaling messages om he UEs and o wa d hem o he
i ualized MME. Each eNB keeps a use inac i i y ime , which has an expi a ion ime
TI
, o each
a ached UE wi hin i s co e age a ea. Using his ime , he eNB de ec s he use s’ inac i i y (i.e., a
use does no pe o m any da a ansmission o e a pe iod o leng h
TI
). I he ime expi es, he eNB
igge s a Se ice Release (SRR) p ocedu e o elease he bea e [5].
3.4. The Vi ualized Mobili y Managemen En i y
The i ualized MME ( MME) is he main con ol plane en i y o he ne wo k. I main ains
he mobili y s a e o he UE and is esponsible o he bea e s’ managemen . Fo simplici y easons,
we assume ha he MME is colloca ed wi h he SGW and he PGW a a cen alized da a cen e o
he ne wo k.
We conside ha he MME is i ualized ollowing he NFV pa adigm, and in pa icula , we adop
he 1:N mapping a chi ec u al op ion. Thus, ollowing he implemen a ion in [
9
,
21
], he MME is spli
in o h ee ie s: FE, MME SL and SDB. We u he assume ha when an MME SL ins ance inishes
p ocessing a con ol plane message, i sa es he ansac ion s a e in o he SDB. When a subsequen
eques a i es a an MME SL ins ance, i i s ga he s he ansac ion s a e om he da abase om
which o con inue. This di e s om he MME implemen a ion in [
21
], and i allows ully s a eless
MME SLs. Di e en messages o he same p ocedu e o he same use can be p ocessed by di e en
MME SL ins ances. The e o e, he numbe o MME SL ins ances, deno ed as
mSL
, can g ow wi hou
a ec ing in-session use s.
As in [
22
], we assume ha all ie s can be eplica ed wi hin limi s. When he p ocessing capaci y
assigned o he MME canno wi hs and he cu en load, a new MME SL ins ance mus be ins an ia ed,
and a new p ocesso is added o he p ocessing esou ces pool.
Fo simplici y, we will assume ha e e y p ocesso in he da a cen e acili y p o ides he same
compu a ional powe , and he SDB ollows a sha ed-e e y hing a chi ec u e, which eases he scale-ou
o scale-in. Then, he scaling o he SDB can be done on-demand, bu wi h ce ain cons ain s [
22
]. These
cons ain s a e due o he sha ed esou ces by mul iple p ocesso s, which may esul in choking o he
bandwid h o simul aneous memo y access o di icul synch oniza ion mechanisms o main ain a
sha ed consis en s a e. Fu he mo e, we will assume each ie has enough memo y esou ces o s o e
an unlimi ed amoun o eques s ha a i e a he ie , while hey a e wai ing o be p ocessed.
4. T a ic Models
We assume h ee classes o a ic: Mobile B oadband a ic (MBB), massi e M2M (mM2M) and
low la ency M2M (lM2M) a ic. In he nex subsec ions, we explain each one.
Senso s 2016,16, 1338 9 o 25
4.1. Mobile B oadband T a ic Model
We adop he compound da a a ic and use beha io model p oposed in [
32
] o MBB UEs.
This a ic model conside s h ee ypes o applica ions, namely:
1. Web b owsing.
2. HTTP p og essi e ideo (e.g., YouTube).
3. Video calling (e.g., Skype se ice).
which a e edesigned o gene a e he da a a es p edic ed o u u e mobile ne wo ks in he METIS
p ojec [
33
]. Then, he mean da a a es pe use used a e highe han he cu en demand. Fo web
b owsing, he mean da a a e pe use depends on he numbe o web pages isi ed, he main objec
size o each web and he numbe o embedded objec s and hei sizes. Fo HTTP p og essi e ideo,
he mean da a a e pe use depends on he numbe o downloaded ideo clips, he size o each
ideo and he ideo encoding a e. While o ideo calling, he mean da a a e pe use depends
on he cons an bi a e o he call. Consequen ly, om [
32
], he mean da a a es pe use ob ained
a e 233.28 kbps, 5.25 Mbps and 142.07 kbps o web b owsing, HTTP p og essi e ideo and ideo
calling, espec i ely.
Mos o he signaling wo kload gene a ed by UEs depends on hei ac i i y and hei a ic
cha ac e is ics [
32
]. He e, we only conside hose LTE con ol p ocedu es ha gene a e he mos
signaling load on he MME [
10
]. In pa icula , we conside Se ice Reques (SR), Se ice Release (SRR)
and X2-based Hando e (HO) p ocedu es. Acco ding o [
32
], o each o hese p ocedu es, he MME
espec i ely will p ocess 3, 3 and 2 con ol messages.
Le
NMBB
be he numbe o MBB UEs and
λSR
,
λSRR
and
λHO
be he mean gene a ion a e pe
UE o he SR, SRR and HO con ol p ocedu es, espec i ely. Le
λMBB
be de ined as he mean a i al
a e o con ol messages p ocessed by he MME ha a e gene a ed by MBB UEs. Then, i can be
compu ed as:
λMBB =NMBB ·(3·λSR +3·λSRR +2·λHO)(2)
4.2. Machine o Machine T a ic Model
We conside wo ypes o M2M de ices: massi e M2M (mM2M) de ices and low la ency M2M
(lM2M) de ices. Le
ph
, whe e
h∈{mM2M,lM2M}
, deno e he pe cen age o each ype o M2M
de ices, and le
be he a io o M2M de ices pe MBB UE. Then, he numbe o M2M de ices o each
ype, Nh, is gi en by:
Nh=NMBB · ·ph(3)
Fo bo h ypes o M2M de ices, we assume a a ic model based on he ansmission o epo s.
Pa icula ly, he a ic model o any M2M de ice has wo possible s a es: ac i e and ala m (see Figu e 4).
M2M de ices gene a e small da a packe s ollowing a Poisson p ocess in bo h s a es. In he ac i e s a e,
packe s a e gene a ed in equen ly, wi h a mean a e
λac i e
. When an e en happens, a pe cen age o
massi e o low la ency M2M de ices, deno ed by
ah
whe e
h∈{mM2M,lM2M}
, change o he ala m
s a e. Du ing he ala m s a e, M2M de ices gene a e packe s mo e equen ly, wi h a mean a e deno ed
as
λala m
. We assume ha he e en has a ixed ime du a ion
e
, and he p obabili y ha in a ce ain
ins an o ime an M2M de ice is in he ala m s a e is deno ed as
pe
h
, whe e
h∈{mM2M,lM2M}
. A e
he e en , all de ices in he ala m s a e e u n o he ac i e s a e.
Massi e M2M de ice e en s occu a ce ain ins an s in ime caused by a la ge numbe o mM2M
de ices synch onizing hei epo ansmissions (see Sec ion 2.3). Some o hese e en s a e pe iodic,
whe eas o he s a e no . Fo simplici y, in ou model, we assume ha mM2M synch oniza ion e en s
ake place a pe iodic ins an s o ime. On he o he hand, lM2M e en s a e caused by ala m si ua ions;
ha is, hey a e unp edic able. Consequen ly, in ou model, lM2M e en s occu a andom ins an s
in ime.
Senso s 2016,16, 1338 16 o 25
Gi en a a ge a i al a e, he goal o ou MME dimensioning is o de e mine he minimum
numbe o ins ances equi ed a each ie
j
o gua an ee he mean esponse ime budge
Dj
. Since we
conside Jackson’s open queuing ne wo k, he mean esponse ime o he sys em
T
can be compu ed as:
T=∑
j
Tj(16)
whe e
Tj
is he mean esponse ime a each ie
j∈ {FE
,
SL
,
SDB}
. Since we assume ha each ie is
modeled by an M/M/m queue, i holds ha :
Tj=1
µj
+C(mj,ρj)
mj·µj−λj
(17)
whe e
ρj=λj
µj
,
µj
is he se ice a e o one ie ins ance,
λj
is he a ge a i al a e conside ed o
dimensioning a ie
j
(see Table 2),
mj
is he numbe o ins ances o he ie and
C(mj
,
ρj)
is E lang’s C
o mula.
C(mj
,
ρj)
ep esen s he p obabili y ha an a i ing packe has o wai in he queue o he ie
because all o he ins ances a e busy, and i has he ollowing exp ession:
C(mj,ρj) = (mj·ρj)mj
mj!·1
1−ρj
∑mj−1
k=0
(mj·ρj)k
k!+(mj·ρj)mj
mj!·1
1−ρj(18)
The p ocessing imes o he FE, SDB and ou pu in e ace a e cons an . Howe e , he p ocessing
ime o an SL is di e en o each con ol message [
32
]. Consequen ly, he mean se ice ime o an SL,
SL =1
µSL
, will depend on he equency wi h which each ype o con ol p ocedu e occu s. Fo his
eason, SL will be di e en o each scheme conside ed in his wo k.
Le
SRi
,
SRRi
,
HOi
and
MO−CPCEOi
deno e he p ocessing ime o he
i
- h message o he SR,
SRR, HO and MO-CPCEO p ocedu e, espec i ely. The mean se ice imes o an SL o each scheme
conside ed a e summa ized in Table 3.
We pe o m dimensioning o each ie indi idually. The dimensioning p oblem o each ie can
be o mula ed as:
mj=min{Mj:Tj(λj,Mj)≤Dj,Mj∈N}(19)
whe e
Dj
is he a ge mean esponse ime o each ie . Hence,
mj
can be compu ed wi h a simple
i e a i e algo i hm ha inc eases he numbe o ie ins ances un il he condi ion
Tj(λj
,
Mj)≤Dj
is me .
Table 3. SL’s mean se ice ime.
Scheme Mean Se ice Time
Baseline
Scheme (BS) SL =NMBB ·(λSR·( SR1+ SR2+ SR3)+λSRR·( SR1+ SR2+ SR3)+λHO ·( HO1+ HO2))+(λin smoo h
mM2M+λin smoo h
lM2M)· MO−CPCEO1
λMBB+λin smoo h
mM2M+λin smoo h
lM2M
O e dimensioned
Scheme (OS) SL =NMBB ·(λSR·( SR1+ SR2+ SR3)+λSRR·( SR1+ SR2+ SR3)+λHO ·( HO1+ HO2))+(λpeak
mM2M+λin smoo h
lM2M)· MO−CPCEO1
λMBB+λpeak
mM2M+λin smoo h
lM2M
T a ic sepa a ed
Scheme (TS)
MBB SL =NMBB ·(λSR·( SR1+ SR2+ SR3)+λSRR·( SR1+ SR2+ SR3)+λHO ·( HO1+ HO2))
λMBB
mM2M SL = MO−CPCEO1
lM2M SL = MO−CPCEO1
T a ic Shape
Scheme (SS) SL =NMBB ·(λSR·( SR1+ SR2+ SR3)+λSRR·( SR1+ SR2+ SR3)+λHO ·( HO1+ HO2))+(λsmoo hpeak
mM2M+λpeak
lM2M)· MO−CPCEO1
λMBB+λsmoo hpeak
mM2M+λpeak
lM2M

Senso s 2016,16, 1338 17 o 25
7. E alua ion
This sec ion includes he simula ion esul s ob ained o e alua e he ou MME schemes
(Sec ion 5). A e he simula ion se up subsec ion, we compa e he equi ed esou ces and hei
associa ed cos s in he dimensioning subsec ion. La e , we compa e he delay expe ienced by each
a ic class in he MME wi h he ou i ualized schemes.
7.1. Simula ion Se up
Ou e alua ion me hodology includes h ee s eps, namely:
1.
The dimensioning o each ie o he MME model and he es ima ion o he associa ed cos using
he a ge a i al a e as inpu ; his es ima ion is done o a ange o UEs and a gi en a io o
M2M de ices pe UE;
2.
The gene a ion o signaling aces o each a ic class (MBB, mM2M and lM2M); simila ly, his
ace gene a ion is done assuming a speci ic numbe o UEs and a gi en a io o M2M de ices
pe UE;
3. Finally, he simula ion o he MME queuing model using he signaling ace as inpu .
We use he MATLAB Simulink amewo k o simula e he queuing sys em p esen ed in Sec ion 6,
which p o ides he MME esponse ime expe ienced by a con ol plane message. The queuing model
is ed wi h he aces o signaling packe s gene a ed by each a ic class. In he model, he se ice a es
o he FE ie , SL ie , SDB ie and ou pu in e ace a e ex ac ed om [32].
Fo he MO-CPCEO p ocedu e, we ob ained ha he p ocessing ime needed a an SL ins ance is
145.05
µ
s. O he main pa ame e s o he simula ion a e summa ized in Table 4. To a oid excessi ely
long simula ions, we assume a pe iod o ime be ween mM2M e en s equal o 60 s.
Table 4. Pa ame e s’ con igu a ion.
Simula ion Pa ame e s
Simula ion Time 300 s
MBB UEs 636,000
M2M de ices pe MBB UE 10
M2M packe size 200 B [19]
M2M e en du a ion 1 s
FE mean esponse ime budge 1 ms
SL mean esponse ime budge 1 ms
SDB mean esponse ime budge 1 ms
Unweigh ed sliding-a e age smoo h 90 ms
wmM2M0.1
wlM2M0.0033
T a ic Models
Mobile B oadband
(MBB)
λSR 0.0045 pk /s [32]
λSRR 0.0045 pk /s [32]
λHO 0.0012 pk /s [32]
Bo h M2M de ices λac i e 0.0033 pk /s
λala m 0.033 pk /s
Massi e M2M
(mM2M)
Pe cen age o mM2M de ices 90%
E en ’s pe iod 60 s
E en ’s magni ude alues [10, 30, 50, 8] %
E en magni ude alues’ p obabili y mass [5, 60, 20, 15] %
Low la ency
M2M (lM2M)
Pe cen age o lM2M de ices 10%
E en ’s pe iod Unique (a 40 s o he simula ion)
E en ’s magni ude alue 33%
Senso s 2016,16, 1338 18 o 25
To es ima e he sys em unning cos , we conside he Amazon EC2 se ice, wi h he cos s and
con igu a ion de ailed in Table 5. We assume a medium-sized CPU ins ance m3.xla ge wi h an a e age
o 11.38
×
10
9
loa ope a ions pe second [
36
]. We use he p ice o he load balancing se ice p o ided
by Amazon o es ima e he cos o he FE ie . Ou se up also includes he Amazon Au o a da abase [
37
],
which is epo ed o p o ide 10
5
upda es/s ansac ions. The o e all cos includes he pe ins ance
cos , he ime-based en al ee and he da a a ic p ocessed.
Table 5. Cloud se ice con igu a ion and cos calcula ion.
Cos Con igu a ion Calcula ion
Cci ype (k)m3.xla ge ins ance en al (0.266 $/h) 0.266/3600
Ccis o (k)Local s o age (10 GB/mon h) and op imized da a access (0.025 $/h). 10 ·0.10 +0.025/3600
Cci h o (k)Da a sen om he da a cen e , (λ(message/s)·200 (by e/message))
0.000 ($)/GB Fi s GB/mon h
0.090 ($)/GB Up o 10 TB/mon h
0.085 ($)/GB Nex 40 TB/mon h
0.070 ($)/GB Nex 100 TB/mon h
0.050 ($)/GB Nex 350 TB/mon h
Cdb ype (k)Au o a db. 3.8xla ge ins ance (4.64 $/h) 4.64/3600
Cdbs o (k)0.1 $ pe GB/mon h, o a o al da abase size o NU·1 KB (0.1 ·NU·1024 ·λ/109)/2,628,000
Cdb h o (k)0.2 $ pe million ansac ions/mon h 0.2 ·λ/106
Cb ype (k)Se ice ee o 0.025 $/mon h 0.025/2,628,000
Cb h o (k)0.008$ pe GB se iced, supposing Osize =200 by es λ·0.008 ·200/109
7.2. Resul s
7.2.1. Dimensioning
We ca ied ou he dimensioning a each ie o he MME e sus
NMBB
o all o he schemes
conside ed (see Figu e 10) by using he heo e ical amewo k desc ibed in Sec ion 6and assuming
Dj=1ms ∀j.
No ably, o simpli y he compa isons in he T a ic sepa a ed Scheme (TS), we se he same
esponse ime budge o MBB and lM2M a ic classes; howe e , his scheme enables one o se
di e en budge s o each class. No e addi ionally ha , o he TS scheme, he equi ed numbe o
SL ins ances
mSL
is he sum o he equi ed numbe o SL ins ances o each ype o a ic, which a e
depic ed in Figu e 10d. Addi ionally, we compu ed he cos pe hou o each scheme conside ed
(see Figu e 11).
As was expec ed, he O e dimensioned Scheme (OS) demands he g ea es amoun o esou ces,
being he mos expensi e scheme. Con e sely, he Baseline Scheme (BS) is he leas expensi e one.
The T a ic sepa a ed Scheme (TS) and he a ic Shape Scheme (SS), which ha e a simila cos , achie e
a no iceable educ ion in cos in compa ison wi h OS. This is mainly hanks o he isola ion be ween
a ic ypes in he TS case and he limi a ion imposed by he a ic shape on he mM2M a ic
a i al a e in he SS case. In such cases, he dimensioning a he SL and SDB ie s can be pe o med
wi hou conside ing
λpeak
mM2M
, which is a ound 12.47- imes g ea e han
λpeak
lM2M
in ou expe imen al se up.
Howe e , bo h he TS and SS schemes a e designed o sa is y he delay cons ain o lM2M a ic.
Senso s 2016,16, 1338 19 o 25
0 0.5 1 1.5 2
x 106
1
1.5
2
2.5
3
3.5
4
Numbe o MBB UEs
Numbe o FE ins ances
BS
OS
TS
SS
(a)
0 0.5 1 1.5 2
x 106
1
1.5
2
2.5
3
3.5
4
Numbe o MBB UEs
Numbe o SDB ins ances
BS
OS
TS
SS
(b)
0 0.5 1 1.5 2
x 106
0
10
20
30
40
50
Numbe o MBB UEs
Numbe o SL ins ances
BS
OS
TS (To al)
SS
(c)
0 0.5 1 1.5 2
x 106
0
2
4
6
8
10
Numbe o MBB UEs
Numbe o SL ins ances
# SL o :
MBB
mM2M
lM2M
(d)
Figu e 10.
Dimensioning a each ie o he MME model ( en M2M de ices pe MBB UE). (
a
) FE
dimensioning; (b) SDB dimensioning; (c) SL dimensioning; (d) De ailed SL TS dimensioning.
0 0.5 1 1.5 2
x 106
0
100
200
300
400
500
600
700
Numbe o MBB UEs
Dimensioning cos s pe hou ($/h)
BS
OS ( o al)
TS
SS
Figu e 11. Dimensioning cos s compa ison pe e alua ed scheme.
Senso s 2016,16, 1338 20 o 25
Addi ionally, Table 6summa izes he es ima ion o he memo y consump ion o each ie o
he MME. I assumes he same numbe o MBB UEs and M2M de ices han he simula ion se up.
We use he numbe o SL ins ances om Figu e 10c. We u ilize 16,545 packe s queued in he sys em.
This numbe is he wo s case o packe s queued o he a ic sepa a ed scheme p oposed, calcula ed
om he esul s ob ained in he nex subsec ion.
Table 6. Memo y consump ion es ima ion (UE con ex ex ac ed om [38,39]).
Elemen Memo y Consump ion
Sample Scena io
(NMBB =636, 000 MBB UEs
NM2M=10 ·NM BB )
S a e
da abase UE con ex : 264 B/UE 264 B/UE ·(NMBB +NM2M)= 1846 MB
Se ice logic
Ope a ing Sys em ROM: 1000 MB/ins ance
Ope a ing Sys em RAM: 400 MB/ins ance [40]
UE con ex : 264 B/UE
Packe size: 200 B
ROM: 1000 MB ·7 ins ances = 7000 MB
RAM: ( 400 MB + 264 B/UE ) ·7 ins ances
+ 16,545 packe s ·200 B/packe = 2803 MB
7.2.2. Delay
We s udied he esponse ime expe ienced by a con ol packe a he SL’s ie o he MME o all
o he schemes (see Figu e 12). Addi ionally, we compu ed he CDF o he o e all sys em esponse
ime, which is he sum o he delay expe ienced by a packe a each ie o he MME (see Figu e 13).
To ha end, we gene a ed a signaling ace o 636,000 MBB UEs and 300 s o du a ion. The ace
includes he h ee conside ed classes. Table 7summa izes he esul s o his poin . The numbe o UEs
is selec ed such ha he p ocessing capaci y o he OS scheme and he lM2M SL pool in he TS scheme
a e abou o equi e an addi ional SL ins ance o sa is y
DSL
. Howe e , wi h his numbe o use s,
he emaining schemes do no expe ience he same si ua ion. The same signaling aces we e used
o he ou schemes conside ed. The esponse ime esul s a e il e ed wi h a simple 90-ms mo ing
a e age. The esul s show ha esponse ime is highe han he a ge mean esponse ime a he SL ie
(
DSL = 1ms
) du ing he mM2M ala m e en s o he BS case (Figu e 12a). Tha is because he SL ie is
unde -dimensioned o suppo he mM2M a ic peaks. Consequen ly, in such si ua ions, he mM2M
a ic migh delay he o he a ic ypes, which may be delay-sensi i e, such as lM2M
(see Figu e 13a)
.
On he con a y, o he OS case, he esponse ime a he SL ie is all o he ime below
DSL
since he
sys em is o e dimensioned (Figu e 12b).
Fo he SS app oach, he SL ie esponse ime always mee s he condi ion
TSL ≤DSL
(
Figu e 12c
).
Tha is because he mM2M a ic peaks a e limi ed by he a ic shape . Mo eo e , du ing he lM2M
a ic peaks, he sys em akes ad an age o he mul iplexing gain. In he TS case, he lM2M pool o
he SL ie also mee s he condi ion
TSL ≤DSL
du ing he lM2M a ic peak (see Figu e 12 ). On he
con a y, he mM2M pool o he SL ie exceeds by se e al o de s o magni ude he esponse ime
budge du ing and a e he mM2M a ic peaks (see Figu e 12e), as mode a e smoo hing o he peaks
is applied. The MBB pool o he SL ie also mee s he esponse ime budge condi ion (see Figu e 12d).
Recall ha , wi h he selec ed numbe o UEs, only he OS (Figu e 12b) and he lM2M class in he TS
scheme (Figu e 12 ) ha e a p ocessing load close o he dimensioned capaci y.
Senso s 2016,16, 1338 21 o 25
0 100 200 300
0
0.5
1
1.5
2
Time (s)
SL delay (s)
mM2M peaks
(a)
0 100 200 300
0
0.2
0.4
0.6
0.8
1x 10−3
Time (s)
SL delay (s)
mM2M peaks
(b)
0 100 200 300
1
1.1
1.2
1.3
1.4
1.5
1.6x 10 4
Time (s)
SL delay (s)
mM2M peaks
lM2M peaks
(c)
0 100 200 300
1
2
x 10−4
Time (s)
SL delay (s)
(d)
0 100 200 300
0
0.5
1
1.5
2
Time (s)
SL delay (s)
mM2M peaks
(e)
0 100 200 300
0
0.2
0.4
0.6
0.8
1
1.2x 10 3
Time (s)
SL delay (s)
lM2M peaks
( )
Figu e 12.
SLs’ il e ed p ocessing ime o each scheme ( en M2M de ices pe MBB UE). (
a
) Baseline
scheme; (
b
) O e dimensioned scheme; (
c
) T a ic shape scheme; (
d
) T a ic sepa a ed scheme: MBB
class; (e) T a ic sepa a ed scheme: mM2M class; ( ) T a ic sepa a ed scheme: lM2M class.

Senso s 2016,16, 1338 22 o 25
Table 7. MME model dimensioning a he simula ion poin .
Scheme
Tie Numbe o FE Ins ances Numbe o SL Ins ances Numbe o SDB Ins ances
BS 1 5 1
OS 2 13 2
TS 2
MBB 2
mM2M 4
lM2M 1
1
SS 2 7 1
10 410 310 210 1100101
0
0.2
0.4
0.6
0.8
1
Time (s)
Empi ical CDF
MBB
mM2M
lM2M
(a)
10 410 3
0
0.2
0.4
0.6
0.8
1
Time (s)
Empi ical CDF
MBB
mM2M
lM2M
(b)
10 410 310 210 1100101
0
0.2
0.4
0.6
0.8
1
Time (s)
Empi ical CDF
MBB
mM2M
lM2M
(c)
10 410 310 210 1100101
0
0.2
0.4
0.6
0.8
1
Time (s)
Empi ical CDF
MBB
mM2M
lM2M
(d)
Figu e 13.
CDF o he il e ed MME delay o each scheme ( en M2M de ices pe MBB UE). (
a
) Baseline
scheme; (b) O e dimensioned scheme; (c) T a ic sepa a ed scheme; (d) T a ic shape scheme.
8. Conclusions
In his pape , we p opose wo designs o a i ualized MME, which aim a acili a ing IoT suppo
in 5G sys ems. The i s p oposed design pa ially sepa a es he p ocessing esou ces de o ed o each
a ic class; while he second design includes a ic shaping o con ol he a ic o each class.
We ha e conside ed h ee a ic classes: MBB, massi e M2M and low la ency M2M. In M2M
communica ions, we ha e included M2M e en s o analyze he pe o mance o he i ualized MME.
M2M e en s conside ed a e caused by he epo synch oniza ion o massi e M2M de ices and ala m
e en s o low la ency M2M de ices. Addi ionally, we assume he use o he CPCEO p ocedu e o
ans e hese epo s om M2M de ices.
Senso s 2016,16, 1338 23 o 25
We ha e compa ed ou p oposed designs wi h wo o he schemes: (i) a baseline i ualized MME
design, which does no apply such esou ce sepa a ion; (ii) an o e dimensioned i ualized MME.
The epo ed compa isons include: (i) dimensioning o he equi ed esou ces; (ii) es ima ion o
he cos s based on he model o Amazon EC2; (iii) he e alua ion o he esponse ime o he i ualized
MME schemes o each a ic class.
A e he conduc ed simula ions, he esul s show ha ou p oposed schemes p o ide much
lowe cos s han he o e dimensioned scheme while hey sa is y he exigen delay equi emen s o
MBB and low la ency M2M communica ions. Fu he mo e, he compa ison o he a ic sepa a ion
scheme and he a ic shape scheme shows ha he mul iplexing gain o he la e p o ides bene i s
in e ms o la ency educ ion. Howe e , he a ic sepa a ion scheme enables one o ha e di e en
delay equi emen s o each a ic class, and addi ionally, i can isola e hei pe o mance.
In addi ion o he abo e esul s, we iden i y he ollowing ad an ages o he conside ed solu ion:
(i) CPCEO p ocedu es mi iga e he signaling explosion gene a ed by a huge numbe o M2M connec ed
de ices. Howe e , i inc eases he p ocessing load on he con ol plane o he EPC. (ii) NFV inc eases
he scalabili y o he ne wo k o deal wi h such load inc ease o signaling. (iii) The usage o ou
p oposed schemes u he op imize he cos s while sa is ying he delay demands.
Howe e , ou solu ion has also he ollowing implica ions: (i) he dimensioning o he esou ces
is mo e complex; (ii) CPCEO p ocedu es use he anspo ne wo k o he con ol plane o send
da a packe s, which imposes an addi ional load. Fu he mo e, ega ding he bo lenecks o he
p oposed schemes, he s a e da abase ie is c i ical. This is due o he s a e da abase scaling wi h
ce ain cons ain s, and ou solu ion makes hea y use o i . Mo eo e , he se ice logic ie design is
de e minan , as i conside ably a ec s he pe o mance o each a ic class.
Acknowledgmen s:
This wo k is pa ially suppo ed by he Spanish Minis y o Economy and Compe i i eness
and he Eu opean Regional De elopmen Fund (P ojec TIN2013-46223-P) and he Spanish Minis y o Educa ion,
Cul u e and Spo (FPU G an 13/04833).
Au ho Con ibu ions:
The wo k p esen ed in his pape is a collabo a i e de elopmen by all o he au ho s.
Juan Manuel Lopez-Sole led he esea ch eam. Pablo Ameigei as de ined he esea ch heme and concei ed o he
sys em. Pila And es-Maldonado and Jona han P ados-Ga zon de eloped he sys em and conduc ed expe imen s.
Juan Jose Ramos-Munoz designed and de eloped he cos s es ima ion. All o he au ho s pa icipa ed in he
w i ing o he pape .
Con lic s o In e es : The au ho s decla e no con lic o in e es .
Re e ences
1.
Holle , J.; Tsia sis, V.; Mulligan, C.; A esand, S.; Ka nouskos, S.; Boyle, D. F om Machine- o-Machine o he
In e ne o Things: In oduc ion o a New Age o In elligence; Else ie : Ox o d, UK, 2014.
2.
E icsson. Ope a o Oppo uni ies in he In e ne o Things—Ge ing Close o he Vision o Mo e Than 50
Billion Connec ed De ices. E icsson Re iew. A ailable online: h p://www.e icsson.com/news/110211_
edcp_244188811_c (accessed on 26 Ap il 2016).
3.
Sha ia mada i, H.; Ra asuk, R.; I aji, S.; Laya, A.; Taleb, T.; Jän i, R.; Ghosh, A. Machine- ype
communica ions: Cu en s a us and u u e pe spec i es owa d 5G sys ems. IEEE Commun. Mag.
2015
,
53, 10–17.
4.
Taleb, T.; Kunz, A. Machine ype communica ions in 3GPP ne wo ks: Po en ial, challenges, and solu ions.
IEEE Commun. Mag. 2012,50, 178–184.
5.
3GPP. TS 23.401 Gene al Packe Radio Se ice (GPRS) Enhancemen s o E ol ed Uni e sal Te es ial Radio Access
Ne wo k (E-UTRAN) Access (V13.6.1); Technical Speci ica ion: Valbonne, F ance, 2016.
6.
Nokia Siemens Ne wo ks. Signaling Is G owing 50% Fas e Than Da a T a ic.
Whi e Pape .
A ailable online: h p://www.ne wo ks.nokia.com/sys em/ iles%20/documen /signaling_whi epape _
online_ e sion_ inal.pd (accessed on 26 Ap il 2016).
7.
Agyapong, P.K.; Iwamu a, M.; S aehle, D.; Kiess, W.; Benjebbou , A. Design conside a ions o a 5G ne wo k
a chi ec u e. IEEE Commun. Mag. 2014,52, 65–75.
Senso s 2016,16, 1338 24 o 25
8.
Hawilo, H.; Shami, A.; Mi ahmadi, M.; Asal, R. NFV: S a e o he a , challenges, and implemen a ion in
nex gene a ion mobile ne wo ks ( EPC). IEEE Ne w. Mag. 2014,28, 18–26.
9.
Taleb, T.; Co ici, M.; Pa ada, C.; Jamako ic, A.; Ru ino, S.; Ka agiannis, G.; Magedanz, T. EASE: EPC as a
se ice o ease mobile co e ne wo k deploymen o e cloud. IEEE Ne w. Mag. 2015,29, 78–88.
10.
Hi schman, B.; Meh a, P.; Ramia, K.; Rajan, A.; Dylag, E.; Singh, A.; Mcdonald, M. High-pe o mance
e ol ed packe co e signaling and bea e p ocessing on gene al-pu pose p ocesso s. IEEE Ne w. Mag.
2015
,
29, 6–14.
11.
Ameigei as, P.; Ramos-munoz, J.J.; Schumache , L.; P ados-Ga zon, J.; Na a o-O iz, J.; Lopez-sole , J.M.
Link-le el access cloud a chi ec u e design based on SDN o 5G ne wo ks. IEEE Ne w. Mag.
2015
,29, 24–31.
12. 5G-PPP. 5G Empowe ing Ve ical Indus ies; B ochu e: Heidelbe g, Ge many, 2016.
13.
Tullbe g, H.; Popo ski, P.; Gozal ez-Se ano, D.; Fe l, P.; Li, Z.; Höglund, A.; Uusi alo, M.A.;
D os e, H.;
Bulakci, Ö.; Eichinge , J.; e al. METIS sys em concep : The shape o 5G o come. A ailable online:
h ps://www.me is2020.com/documen s/publica ions/ (accessed on 19 Augus 2016).
14.
Ra asuk, R.; P asad, A.; Li, Z.; Ghosh, A.; Uusi alo, M.A. Recen ad ancemen s in M2M communica ions
in 4G ne wo ks and e olu ion owa ds 5G. In P oceedings o he 2015 18 h In e na ional Con e ence on
In elligence in Nex Gene a ion Ne wo ks (ICIN), Pa is, F ance, 17–19 Feb u ay 2015; pp. 52–57.
15.
3GPP. TR 22.368 Se ice Requi emen s o Machine-Type Communica ions (MTC) (V13.1.0); Technical Repo :
Valbonne, F ance, 2014.
16.
And es-Maldonado, P.; Ameigei as, P.; P ados-Ga zon, J.; Ramos-Munoz, J.J.; Lopez-Sole , J.M. Reduced
M2M signaling communica ions in 3GPP LTE and u u e 5G cellula ne wo ks. In P oceedings o he 2016
Wi eless Days (WD), Toulouse, F ance, 23–25 Ma ch 2016; pp. 1–3.
17.
3GPP. TR 23.887 S udy on Machine-Type Communica ions (MTC) and O he Mobile Da a Applica ions
Communica ions Enhancemen s (V12.0.0); Technical Repo : Valbonne, F ance, 2013.
18.
3GPP. TR 36.802 E ol ed Uni e sal Te es ial Radio Access (E-UTRA); NB-IOT; Technical Repo o BS and
UE Radio T ansmission and Recep ion (V1.0.0); Technical Repo : Valbonne, F ance, 2016.
19.
3GPP. TR 45.820 Technical Speci ica ion G oup GSM/EDGE Radio Access Ne wo k; Cellula Sys em Suppo o
Ul a-Low Complexi y and Low Th oughpu In e ne o Things (CIoT) (V13.1.0); Technical Repo : Valbonne,
F ance, 2015.
20.
Omnes, N.; Bouillon, M.; F omen oux, G.; G and, O.L. A p og ammable and i ualized ne wo k IT
in as uc u e o he in e ne o hings: How can NFV SDN help o acing he upcoming challenges.
In P oceedings o he 2015 18 h In e na ional Con e ence on In elligence in Nex Gene a ion Ne wo ks
(ICIN), Pa is, F ance, 17–19 Feb ua y 2015; pp. 64–69.
21.
Takano, Y.; Khan, A.; Tamu a, M.; Iwashina, S.; Shimizu, T. Vi ualiza ion-based scaling me hods o s a e ul
cellula ne wo k nodes using elas ic co e a chi ec u e. In P oceedings o he 2014 IEEE 6 h In e na ional
Con e ence on Cloud Compu ing Technology and Science (CloudCom), Singapo e, 15–18 Decembe 2014;
pp. 204–209.
22.
U gaonka , B.; Shenoy, P.; Chand a, A.; Goyal, P.; Wood, T. Agile dynamic p o isioning o mul i- ie in e ne
applica ions. ACM T ans. Au on. Adap . Sys . 2008,3, 1–39.
23.
U gaonka , B.; Paci ici, G.; Shenoy, P.; Sp ei ze , M.; Tan awi, A. An analy ical model o mul i- ie in e ne
se ices and i s applica ions. SIGMETRICS Pe o m. E al. Re . 2005,33, 291–302.
24.
Huang, D.; He, B.; Miao, C. A su ey o esou ce managemen in mul i- ie web applica ions. IEEE Commun.
Su . Tu o . 2014,16, 1574–1590.
25.
Amazon Web Se ice. Amazon EC2 Ins ances. A ailable online: h p://www.aws.amazon.com/es/ec2/
ins ance- ypes/ (accessed on 26 Ap il 2016).
26.
Mao, M.; Humph ey, M. A pe o mance s udy on he VM s a up ime in he cloud. In P oceedings o he 2012
IEEE 5 h In e na ional Con e ence on Cloud Compu ing (CLOUD), Honolulu, HI, USA,
24–29 June 2012;
pp. 423–430.
27.
Jupi e Ne wo ks. Machine- o-Machine (M2M) —The Rise o he Machines.
Whi e Pape .
A ailable online:
h p://es.slidesha e.ne /jpocalles/machine- o-machine-whi e-pape (accessed on 19 Augus 2016).
28.
Pö sch, T.; Ma wa , S.N.K.K.; Zaki, Y.; Go g, C. In luence o u u e M2M communica ion on he LTE sys em.
In P oceedings o he 2013 6 h Join IFIP on Wi eless and Mobile Ne wo king Con e ence (WMNC), The Palm,
Dubai, 23–25 Ap il 2013; pp. 1–4.
Senso s 2016,16, 1338 25 o 25
29.
Sha iq, M.Z.; Ji, L.; Liu, A.X.; Pang, J.; Wang, J. La ge-scale measu emen and cha ac e iza ion o cellula
machine- o-machine a ic. IEEE T ans. Ne w. 2013,21, 1960–1973.
30.
3GPP. TD S2-101456 Signalling conges ion in PS co e ne wo k; SA WG2 Tempo a y Documen : Valbonne,
F ance, 2010.
31.
ETSI. Machine- o-Machine Communica ions (M2M); Applicabili y o M2M A chi ec u e o Sma G id Ne wo ks;
Impac o Sma G ids on M2M Pla o m; Technical Repo TR 102 935 (V2.1.1); Technical Repo : Valbonne,
F ance, 2012.
32.
P ados-Ga zon, J.; Ramos-Munoz, J.J.; Ameigei as, P.; And es-Maldonado, P.; Lopez-Sole , J.M. La ency
e alua ion o a i ualized MME. In P oceedings o he 2016 Wi eless Days (WD), Toulouse, F ance,
23–25 Ma ch 2016; pp. 1–3.
33.
Agyapong, P.; B aun, V.; Fallg en M.; Gou aud, A.; Hessle , M.; Jeux, S.; Klein, A.; Ji, L.; Ma ín-Sac is án, D.;
Ma e nia, M.; Simula ion Guidelines (Deli e able D6.1); METIS Technical Repo ; A ailable online: h ps://
www.me is2020.com/wp-con en /uploads/deli e ables/METIS_D6.1_ 1.pd (accessed on 19 Augus 2016).
34.
Amazon Web Se ice. Au o a P icing. A ailable online: h p://www.aws.amazon.com/ ds/au o a/
p icing/ ?nc2 =h_ls (accessed on 26 Ap il 2016).
35.
Vilaplana, J.; Solsona, F.; Teixidó, I.; Ma eo, J.; Abella, F.; Rius, J. A queuing heo y model o cloud compu ing.
J. Supe compu . 2014,69, 492–507.
36.
Iosup, A.; Os e mann, S.; Yigi basi, M.N.; P odan, R.; Fah inge , T.; Epema, D.H. Pe o mance analysis
o cloud compu ing se ices o many- asks scien i ic compu ing. IEEE T ans. Pa allel Dis ib. Sys .
2011
,
22, 931–945.
37.
Amazon Au o a Pe o mance Assessmen ; Technical Repo ; A ailable online: h ps://d0.awss a ic.com/
p oduc -ma ke ing/Au o a/RDS_Au o a_Pe o mance_Assessmen _Benchma king_ 1-2.pd (accessed on
19 Augus 2016).
38. NMC. LTE Iden i ie s. A ailable online: h ps://es.sc ibd.com/doc/311550989/NMC-LTE-Iden i ie s
- 1-0 -pd (accessed on 14 July 2016).
39.
3GPP. TS 33.401 Sys em A chi ec u e E olu ion (SAE); Secu i y A chi ec u e (V13.2.0); Technical Speci ica ion:
Valbonne, F ance, 2016.
40.
Ubun u Documen a ion. Recommended Minimum Sys em Requi emen s. A ailable online:
h ps://help.ubun u.com/communi y/Ins alla ion/Sys emRequi emen s (accessed on 14 July 2016).
c

2016 by he au ho s; licensee MDPI, Basel, Swi ze land. 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
(CC-BY) license (h p://c ea i ecommons.o g/licenses/by/4.0/).