How Should I Slice My Network? A Multi-Service Empirical ... · radio access, through, e.g., cloud...

How Should I Slice My Network A Multi-ServiceEmpirical Evaluation of Resource Sharing Efficiency

Cristina MarquezUniversidad Carlos III Madrid

Spainmcmarquepauc3mes

Marco GramagliaUniversidad Carlos III Madrid

Spainmgramaglituc3mes

Marco FioreCNR-IEIIT

Italymarcofioreieiitcnrit

Albert BanchsUniversidad Carlos III Madrid amp

IMDEA Networks InstituteSpain

banchsituc3mes

Xavier Costa-PerezNEC Laboratories Europe

Germanyxaviercostaneclabeu

ABSTRACTBy providing especially tailored instances of a virtual net-work network slicing allows for a strong specialization of theoffered services on the same shared infrastructure Networkslicing has profound implications on resource managementas it entails an inherent trade-off between (i) the need forfully dedicated resources to support service customizationand (ii) the dynamic resource sharing among services toincrease resource efficiency and cost-effectiveness of the sys-tem In this paper we provide a first investigation of thistrade-off via an empirical study of resource management ef-ficiency in network slicing Building on substantial measure-ment data collected in an operational mobile network (i) wequantify the efficiency gap introduced by non-reconfigurableallocation strategies of different kinds of resources from ra-dio access to the core of the network and (ii) we quantifythe advantages of their dynamic orchestration at differenttimescales Our results provide insights on the achievable ef-ficiency of network slicing architectures their dimensioningand their interplay with resource management algorithms

CCS CONCEPTSbull Networks rarr Mobile networks Network architecturesNetwork performance evaluation Network management

Conferencersquo17 July 2017 Washington DC USAcopy Association for Computing Machinery

This is the authorrsquos version of the work It is posted here for your per-sonal use Not for redistribution The definitive Version of Record waspublished in The 24th Annual International Conference on Mobile Computingand Networking (MobiCom rsquo18) October 29-November 2 2018 New DelhiIndia httpsdoiorg10114532415393241567

KEYWORDSNetwork slicing resource management network efficiency

ACM Reference FormatCristina Marquez Marco Gramaglia Marco Fiore Albert Banchsand Xavier Costa-Perez 2018 How Should I Slice My Network AMulti-Service Empirical Evaluation of Resource Sharing EfficiencyIn The 24th Annual International Conference on Mobile Computingand Networking (MobiCom rsquo18) October 29-November 2 2018 NewDelhi India ACM New York NY USA 16 pages httpsdoiorg10114532415393241567

1 INTRODUCTIONCurrent trends in mobile networks point towards a strongdiversification of services which are characterized by in-creasingly heterogeneous Key Performance Indicator (KPI)and Quality of Service (QoS) requirements This tendencyis driving the design of 5G networks that will eventuallyhave to support eg the Internet of Thing (IoT) with ultra-low rate communication from a massive number of devicesautomotive and tactile applications with millisecond laten-cies industrial communications with extreme reliability andvirtualaugmented reality services with very high data rates

However clear needs for tailored KPI and QoS require-ments are already evident in todayrsquos mobile services whichencompass eg high-quality video streaming machine-typecommunication low-latency mobile gaming jointly withbest effort traffic Unfortunately current mobile network ar-chitectures [33] lack the necessary flexibility to meet the ex-treme requirements imposed by such services This situationis pushing independent initiatives to address the problem3GPP has developed a IoT-specific MAC that co-exists withthe legacy general-purpose MAC layer [1] Network deploy-ments in industrial environments rely on proprietary archi-tectures that ensure reliability levels not attainable with pub-lic mobile networks [17] Google started deploying its own

Dedicated Resources

Q

I

0000 0100 1100 1000

0001 0101 1101 1001

0011 0111 1111 1011

0010 0110 1110 1010

Antenna Spectrum RAN processing

Radio Resource Management

Edge Cloud

Core Cloud

Type A Slice

Type B Slice

Type C Slice

Type D Slice

Type E Slice

Shared Resources

Figure 1 Network slicing types Deeper strategies usededicated resources customized to services across awider portion of the end-to-end network architecture

radio access infrastructure and proprietary transit networkto run its many services under hard QoS guarantees [13]

Network virtualization and slicingWhile their scopeis clearly limited these solutions do show the need for cus-tomized network support even with present-day traffic Theyalso substantiate the well-established vision that several net-work instances each devoted to a specific set of serviceshave co-exist in the same infrastructure in order to satisfythe KPI and QoS requirements of current and future mobileapplications The agenda for 5G networks is to achieve thismainly via network virtualization which evolves the tradi-tional hardbox paradigm into a cloudified architecture wherethe once hardware-based network functions (eg spectrummanagement baseband processing mobility management)are implemented as software Virtual Network Functions(VNFs) running on a general-purpose telco-cloud Networkvirtualization enables the deployment of multiple virtualinstances of the complete network named network slicesSlices are then easily customized by tuning the functionalityand location of VNFs They thus create on top of the physi-cal infrastructure a set of logical networks each tailored toaccommodate fine-tuned Service Level Agreements (SLA)reflecting the needs of different service providers

Network slicing and resourcemanagementNetworkslicing has profound implications on resource managementWhen instantiating a slice an operator needs to allocate suf-ficient computational and communication resources to itsVNFs In some cases these resources may be dedicated be-coming inaccessible to other slices [26] Alternatively smartassignment algorithms can be employed to dynamically allo-cate resources to slices based on the time-varying demandsof tenants [11 15] This grants the flexibility to modify theshare of resources assigned to each tenant multiplexing logi-cal slices into the software or hardware assets while trying toabide by tenant requirements However such algorithms in-troduce additional complexity and may in some cases hinderresource isolation the corresponding guarantees to tenantsandor the ability to deploy fully customized slices

The above shows that there is an inherent trade-off among(i) service customization which favours the deployment ofspecialized slices with tailored functions for each service andpossibly dedicated and guaranteed resources (ii) resourcemanagement efficiency which increases by dynamically shar-ing the resources of the common infrastructure among thedifferent services and slices and (iii) system complexity re-sulting from deploying more dynamic resource allocationmechanisms that provide higher efficiency at the cost of em-ploying elaborate operation and maintenance functions [28]

The above trade-off is fundamentally affected by the strat-egy adopted to implement network slicing as illustrated inFigure 1 In its simplest realization slices are limited to thecore network (type-A slice in Figure 1) the allocation of re-sources to slices only involves cloud resources and mostlybecomes a Virtual Machine (VM) or container resource as-signment problem [14] In this case the level of service cus-tomization granted by slices is low since it is restricted tocore network functions yet high efficiency can be achievedat low complexity as a large portion of the network remainsshared among all services and tenantsMore dependable slicing would offer customized func-

tions possibly involving dedicated resources also at theradio access through eg cloud RAN (C-RAN) paradigmsHere basic radio-access slices allow for tailored MAC-layerscheduling [30] across a large number of antennas (type-B slice) Moving down the protocol stack advanced slicesimplement customized baseband processing (ie encodingand decoding operations) in the Base Band Units (BBUs)possibly providing tenants with a guaranteed bandwidth atthe air interface (type-C slice) These approaches provide theability to customize scheduling strategies but at the sametime they reduce the possibility of radio resource sharingandor increase the system complexity

At fronthaul resource isolation becomes a hardware prob-lem [31] A first case for slicing is one where tenants shareantenna sites but are granted their own dedicated spectrum(type-D slice) we have virtually independent protocol stacksand full isolation and sharing is limited to the physical hard-ware Otherwise tenants may require dedicated end-to-endresources down to the antennas (type-E slice) this resultsinto slices that tell apart full end-to-end virtual networks

In general slicing strategies at the higher network layersprovide a lower level of customization yet they can moreeasily achieve efficient resource sharing without additionalcomplexity Indeed when slicing occurs at high layers (egtype-A) the operator cannot offer full customization but itcan easily employ highly dynamic allocation schemes for thelower layers in contrast achieving such an efficient resourceallocation is much more challenging when considering net-work slicing schemes with stringent customization require-ments (ie strategies involving the lower layers down to

type-E slicing) For instance when all slices have a commonMAC layer an efficient sharing of radio resources is easy yetMAC is not tailored to their different needs conversely ifeach slice implements a different customized MAC protocolit is more difficult to efficiently share radio resources

Contribution of this paper From a system standpointthe technology needed to support the different types of slicesis well understood or even already available For instancethere exist several cloud resource orchestrators for both com-mercial and open-source telco-cloud platforms [23] similarlya variety of solutions have been proposed for the dynamicallocation of resources across network slices [14]However the implications of network slicing in terms of

efficiency of network resource utilization are still not wellunderstood Efficiency intuitively grows as one moves awayfrom the radio access infrastructure (type-E slicing) towardsthe network core (type-A slicing) but we lack any moredetailed characterization of the aforementioned trade-offsbetween customization efficiency and complexity This isan important gap since insights on the efficiency gains innetwork slicing are crucial to take informed decision onresource configuration strategies if efficiency is preservedwith solutions that assign resources to slices more or lessstatically high customization levels can be achieved at areduced complexity however if the price in efficiency is highmore elaborate (and expensive) solutions may be desirableOur aim is to shed light on the trade-offs between cus-

tomization efficiency and complexity in network slicingby evaluating the impact of resource allocation dynamics atdifferent network points Based on our analysis it is thuspossible to determine in which cases the gains in efficiencyare worth the sacrifice in customizationisolation andor theextra complexity Since resource management efficiency innetwork slicing highly depends on the traffic patterns ofdifferent services supported by the various slices we buildon substantial service-level measurement data collected by amajor operator in a production mobile network and

(i) quantify the price paid in efficiency when suitable algo-rithms for dynamic resource allocation are not available andthe operator has to resort to physical network duplication(ii) evaluate the impact of sharing resources at different

locations of the network including the cloudified core thevirtualized radio access or the individual antennas

(iii) outline the benefit of dynamic resource allocationat different timescales ie allowing to reallocate resourcesacross slices with different reconfiguration intervals

To the best of our knowledge this is the first work tacklingthe empirical assessment of network slicing in real-worldnetworks We believe that the insights it provides can beused as rule of thumb to evaluate the solution space forsmart resource assignment algorithms and infrastructuredimensioning For instance our results show that efficiency

Slice a Slice b

Figure 2 Hierarchical mobile network architectureNodes map to different equipment depending on thelevel ℓ and form a hierarchy The mobile traffic of ser-vices in each slice (eg a or b) is increasingly aggre-gated as it flows from radio access to network core

gains are very high in the edge where employing technolo-gies that allow for dynamic resource allocation provides ahigh reward in contrast gains are much reduced in the corewhere complex highly flexible reconfiguration schemes maynot always pay off Mobile network operators should thus beaware that isolating slices at the radio access may have a highcost in terms of efficiency and that network slicing shouldbe combined with solutions for dynamic orchestration ofresources at least at the network edge

2 NETWORK SCENARIO AND METRICSIn the following we expose our network scenario our rep-resentation of the slice QoS requirements and a consistentresource allocation strategy and the metrics we adopt toevaluate the resource sharing performance

21 Network slicing scenarioLet us consider a mobile network providing coverage toa generic geographical region where mobile subscribersconsume a variety of heterogeneous services The operatorowning the infrastructure implements slices s isin S eachdedicated to a different subset of services

We assume that each slice can be implemented accordingto any of the strategies in Figure 1 To capture such a generalscenario we model the mobile network architecture as a hi-erarchy composed by a fixed number of levels (ℓ = 1 L)ordered from the most distributed (ℓ = 1) to the most central-ized (ℓ = L) as illustrated in Figure 2 Every network levelℓ is composed by a set Cℓ of network nodes each serving agiven number of base stations In the two extremes we haveℓ = 1 where network nodes in C1 have a bijective mappingto individual antennas and ℓ = L where CL contains a singlenetwork node controlling all antennas in the whole targetregion In between for 1 lt ℓ lt L the number of networknodes in Cℓ decreases with ℓ whereas that of base stationsserved by each such node increases accordingly Note that

in general a node c isin Cℓ will operate on data flows that areincreasingly aggregated with ℓ which as we will see has asignificant impact on resource managementThis hierarchical representation allows considering a va-

riety of node types along with their associate (possibly vir-tual) network functions At the most distributed level (ℓ = 1)each node runs functions that operate at the antenna leveleg involving spectrum or airtime In intermediate cases(1 lt ℓ lt L) nodes are at first in charge a small number ofantenna sites eg C-RAN datacenters running VNFs suchas dedicated baseband processing or radio resource man-agement As ℓ grows VNFs are pushed further towards thenetwork core into telco-cloud datacenters that tunnel trafficto and from large sets of antenna sites there VNFs cus-tomize VM resources for large traffic volumes associated tothe services delivered by each tenant to subscribers in widegeographical areas In the limit case (ℓ = L) all traffic inthe target region is managed in a fully-centralized fashionat a single datacenter where the operator can tailor cloudresources to the whole demand for the services of a tenantNote that in the case of VNFs this allows to evaluate theimpact of instantiating or moving VNFs at different nodesUltimately the layered network model allows generaliz-

ing our analysis to diverse VNFs by studying the systemperformance at different network levels This also implicitlyaccommodates all of the network slicing strategies outlinedin Figure 1 Slices of type-D and type-E deal with the lowestnetwork layers that are implemented at the antennas hencecorrespond to ℓ = 1 Slices of type-A refer to VNFs operatingat higher network layers that are deployed at centralizedcloud datacenters hence correspond to high values of thenetwork level ℓ Slices of type-B and type-C are concernedwith VNFs for radio access resources which may run at thebase stations (ℓ = 1) in a distributed implementation orat higher architectural levels (1 lt ℓ lt L) in a centralizedC-RAN implementation

Note that we do not require that a single network deploysvirtualization technologies at all network levels Instead bytaking a large number of levels and considering each ofthem in isolation this approach lets us cover a wide rangeof deployment options and provide insights for all of them

22 Slice specificationsNetwork slicing allows the operator to fulfil minimum QoSrequirements requested by each tenant We capture suchrequirements as a slice specification z which is established soas to ensure a sufficient service quality for the slice demandsMore precisely a slice specification involves

(i) Guaranteed time fraction f the operator engages toguarantee that the traffic demand of the slice is fullyserviced during at least a fraction f isin [0 1] of time

(ii) Averaging window length w the operator commitmenton fraction f above is intended on discrete-time de-mands of granularityw with traffic averaged over thedisjoint time windows of durationw

We denote such a slice specification as z = ( f w ) whichbecomes more stringent for higher values of f and smallerw

To ensure compliance with the requirements the operatorshall guarantee that enough resources are allocated to allslices s isin S at every node c isin Cℓ of each network levelℓ Formally the required amount of resources needed tomeet a slice specification z = ( f w ) is computed as followsLet ocs (t ) denote the load offered by slice s at node c andtime t also let ocs (k ) = 1

w

intk ocs (t ) dt be the average load

over window k covering a time interval of the same namewith duration w Let us also denote by r zcs (k ) the amountof resources allocated to slice s at node c during window k According to the above requirements r zcs (k ) has to be setsuch that the following inequality holds

P(r zcs (k ) ge ocs (k )

)ge f (1)

where P (middot) denotes the probability of the argument BasicallyEquation (1) states that the resources allocated should meetthe demand for at least a fraction f of averaging windows

Note that the expression in Equation (1) assumes that theamount resources needed to serve a given slice r zcs (k ) isdirectly proportional to the mobile traffic demand in thatslice ocs (k ) While this clearly holds for some types of re-sources (eg radio) we acknowledge that it may be a strongsimplification in other cases We argue however that it isa reasonable assumption for many practical VNFs More-over this choice allows us to investigate through a unifiedframework different network levels ℓ where resources mapto diverse physical assets (such as spectrum airtime CPUtime computational power or memory) depending on ℓ

23 Resource allocation to slicesIn presence of algorithms that enable a dynamic reconfigu-ration of VNFs the resource allocation can be re-modulatedover time If at some node c one could reallocate resourcesat every averaging window it would be sufficient to assignto a slice s the resources it requires during that window withprobability at least f according to Equation (1)However in practice the periodicity of reconfiguration

is limited by the adopted slicing strategy (see Figure 1) aswell as by the constraints of the underlying technology Forinstance when network slicing is performed at the antennalevel non-negligible times in the order of minutes are neededto turn on and off the radio-frequency front-end and resetthe transport network When dealing with radio resource

Mon Tue Wed Thu Fri Sat Sun0

50

100

150

200

Norm

aliz

ed m

obile

t

raff

ic d

em

and

0 50 100 150 200Normalized mobile traffic demand

0

02

04

06

08

10

CD

F

f=09

Figure 3 Example of resource allocation to a slicewithspecification z = ( f w ) Left time series of the mobiletraffic demand for a slice s dedicated to a popular videostreaming service The time series refers to traffic av-eraged over windows of lengthw = 1 hour recorded ata datacenter c serving a medium-sized city (ℓ = L) dur-ing one reconfiguration interval n (τ = 1 week) RightCDF Fwscn of the demand in the left plot The guaran-teed time fraction is f = 09 hence the minimum re-sources r zcs (n) to meet the service requirements underslice specification z = ( f w ) = (09 1 hour) is the 90thpercentile of the distribution highlighted by the ver-tical line The same value is shown in the left plot as ahorizontal line traffic above it is not guaranteed

management algorithms (ie dynamic spectrum or multi-provider scheduling) re-assignments are constrained by sig-nalling overhead Or in the case of VM orchestration thetimescale is limited by instantiation and migration times [20]

Let us assume that τ ≫ w is the minimum time needed forresource reallocation which we refer to as a reconfigurationperiod Let us denote by n isin T the nth reconfiguration pe-riod within the set T of all the reconfiguration periods thatcompose the whole system observation time also r zcs (n) isthe amount of resources allocated to slice s in node c duringthe reconfiguration period n under specification z Sinceno reassignment is possible within a reconfiguration periodthen r zcs (k ) = r zcs (n) for all averaging windows k withinreconfiguration period n In compliance with Equation (1)the allocation of resources at reconfiguration period n shallbe such that the offered load does not exceed r zcs (n) for atleast a fraction f averaging windows encompassed by nLet Fwscn be the Cumulative Distribution Function (CDF)of the demand for slice s at node c during reconfigurationperiod k averaged over windows of lengthw then the min-imum r zcs (n) that satisfies Equation (1) can be computed asr zcs (n) = (Fwscn )

minus1 ( f ) Figure 3 illustrates this concept1Once we have computed r zcs (n) we can define the amount

of resources that the operator will need to allocate at networklevel ℓ over the entire system observation period as

Rzℓτ =sums isinS

sumc isinCℓ

sumnisinT

τ middot r zcs (n) (2)

1All traffic volumes in the paper are normalizedwith respect to theminimumaverage traffic recorded at a 4G antenna sector in our reference scenarios

The above equation represents the total amount of re-sources needed to meet slice specifications z under the pos-sibility of dynamically re-configuring the allocation withperiodicity τ Note that it can accommodate the special casewhere no reconfiguration is possible at level ℓ by setting τto the total system observation time ie |T | = 1

24 Multiplexing efficiencyEquation (2) provides the total amount of resources that theoperator needs to provision in order to satisfy the commit-ments with all tenants In order to unveil the implicationsof this value we compare it against a perfect sharing bench-mark In perfect sharing the allocated resources correspondto those required when there is no isolation among differentservices hence traffic multiplexing is maximum Formally

Pzℓτ =sumc isinCℓ

sumnisinT

τ middot r zc (n) (3)

where r zc (n) denotes the resources needed to accommodatethe traffic demand at node c during reconfiguration period naggregated over all slices For the sake of fairness the samespecification z = ( f w ) assumed for individual slices areenforced in the benchmark provided by Equation (3) Thusr zc (n) = (Fwcn )

minus1 ( f ) where Fwcn is the CDF of the total de-mand for mobile data traffic at node c during reconfigurationperiod n averaged over windows of lengthw

Taking the above benchmark we define the multiplexingefficiency as the ratio between the resources required withnetwork slicing and those needed under perfect sharing ie

Ezℓτ = Rzℓτ P

zℓτ (4)

Equation (4) refers to network level ℓ resource reconfigura-tion intervals of duration τ and slice specification z

In summary Ezℓτ quantifies the efficiency of the network

slicing paradigm in terms of resource management as Ezℓτ

approaches 1 the total amount of slice-isolated resourcestend to that assured by a perfect sharing Indeed with perfectsharing we can allocate resources at a given level accordingto the total peak demand over the reconfiguration periodwhile with network slicing we need to allocate resourcesaccording to the peak demand at each slice which becomesinefficient when such peaks occur at different windows Fig-ure 4 illustrates the intuition behind multiplexing efficiencywith an example

3 CASE STUDIESWe evaluate the efficiency of resource allocation in a slicednetwork by considering two realistic case studies in modernmetropolitan-scale mobile networks As mentioned in theintroduction todayrsquos mobile services already offer a varietyof requirements that makes it meaningful to investigate theimpact of slice isolation on resource management

100

200

300

400

500

600

Nor

mal

ized

mob

iletra

ffic

dem

and

100

200

300

400

500

600

Nor

mal

ized

mob

iletra

ffic

dem

and

f=09

Figure 4 Example of multiplexing efficiency Lefttime series of the mobile traffic demands for a set S offive slices s observed at a single datacenter c servingone medium-sized city (ℓ = L) during one reconfigu-ration interval n (τ = 1 week) The slice specificationz = ( f w ) = (09 1 hour) commits the operator to servefor each slice at least the traffic volumes highlightedby the grey horizontal lines (computed for each timeseries as in Figure 3) The sum value of such lines inthick gold denotes

sums isinS τ middotr

zcs (n) aswe are looking at a

single node c and one specific reconfiguration intervaln this is the traffic volume that provides the needed re-sources according to Equation (2) Right time series ofthe traffic demand aggregated over all services for thesame set of slices By applying an identical slice spec-ification we get the equivalent traffic volume r zc (n) tobe served under perfect sharing as per Equation (3)this is highlighted by the horizontal thick gold lineThe multiplexing efficiency is the ratio between thevalues highlighted by the thick gold lines on the rightand left plots In this toy example the two values areclose hence resource isolation is efficient In practi-cal scenarios (Section 41) we find major differencesbetween network slicing and perfect sharing and re-source isolation proves highly inefficient

Our two reference urban regions are a large metropolis ofseveral millions of inhabitants and a typical medium-sizedcity with a population of around 500000 both situated inEurope Service-level measurement data was collected in thetarget areas by a major operator with a national market shareof around 30 We leverage these real-world traffic demandsto define network slices Details are in Section 31

On top of this we model the hierarchical network infras-tructures in the target regions by assuming that the operatordeploys level-ℓ nodes so as to balance the offered load amongthem This is discussed in Section 32

31 Mobile service demandsThe real-world demands generated by individual mobile ser-vices in the two reference regions were collected by the oper-ator during three months in late 2016 The information wasgathered by monitoring individual IP data sessions over theGPRS Tunneling Protocol User plane (GTP-U) and running

0 10 20 30 40Service rank

0

5

10

15

Perc

enta

ge o

f tra

ffic downlink

uplink

0

5

10

15

Perc

enta

ge o

f tra

ffic downlink

uplink

Figure 5 Percentage of the mobile traffic generated bythe each service in our study The fraction of downlinkand uplink traffic is denoted by different colors Leftlarge metropolis Right medium-sized city

10 5 10 2 10 1 100 101

Normalized mobile traffic demand0

01

02

03PDF

10 5 10 2 10 1 100 101

01

02

03PDF

Figure 6 PDF of traffic demands across antenna sec-tors Left large metropolis Right medium-sized city

Deep Packet Inspection (DPI) and proprietary fingerprintingalgorithms to infer the mobile service associated to each2G3G4G data session The data was aggregated geograph-ically (per antenna sector) and temporally (over 5-minutetime intervals) by the operator so as to make the data non-personal and to preserve user privacy all operations werecarried out within the operator premises under control ofthe local Data Privacy Officer (DPO) and in compliance withapplicable regulationsThe resulting measurement data describe downlink and

uplink traffic for hundreds of prominent mobile services con-sumed in the target regions Building on such informationwe define potential slices by identifying mobile services thatmeet two requirements (i) they generate a substantial of-fered load (above 01 of the total network traffic) sufficientto justify the creation of a dedicated network slice and (ii)they entail clearly distinguishable KPIs and QoS require-ments We identify 38 services that meet the criteria aboveand associate them to a different network slice each

Our choice of services represents well the heterogeneousnature of todayrsquosmobile traffic It encompassesmany popularservices such as YouTube Netflix Snapchat Pokemon GoFacebook or Instagram and covers a wide range of classeswith diverse network requirements including mobile broad-band (eg long-lived and short-lived video streaming) low-latency (eg gaming messaging) and best effort (eg webbrowsing social media) We consider such service classesas representative forerunners of those expected for 5G ser-vices [4] Figure 5 provides basic information on our selec-tion of services It outlines the downlink-dominated highlyskewed traffic split among the services the percent trafficcan differ of more than two orders of magnitude

Figure 7 Antenna deployments in the target regionsLeft large metropolis Right medium-sized city

A strong diversity also emerges in the way the selectedservices are consumed across the geographical space withinthe two urban regions Figure 6 portrays the Probability Den-sity Function (PDF) of the total offered load at individualantenna sectors which again spans several orders of mag-nitude The main cause of heterogeneity is the radio accesstechnology as our measurement data captures 2G 3G and4G access it is natural that 4G antennas accommodate muchlarger fractions of the demand and generate the rightmostbell-shaped lob of the distributions Still 10-time differencesin the traffic volume appear even across 4G antenna sectorsimplying substantial location-based demand variability

32 Hierarchical network structureThe deployment of antennas in the target regions is illus-trated in Figure 7 which highlights the different scales of thetwo case studies in terms of geographical span and densityof user populations and thus of the network infrastructuresneeded to support the local mobile service demands Whilewe do not have information on the architecture of the mobilenetworks beyond the radio access we model the hierarchicalstructure exemplified in Figure 2 after current proposals forcloudified network slicing [24] as followsAt the generic level ℓ the operator deploys a number

Nℓ = |Cℓ | of nodes each responsible for a subset of the an-tenna sites at the radio access level Every node will thusrun VNFs (whose nature will depend on ℓ) on the mobiledata traffic incoming from or outgoing to its associated an-tennas We assume that the operator deploys generic level-ℓnodes and links based on two criteria (i) the offered loadshould be similar at all nodes and (ii) the subset of antennasassociated to a same node shall be geographically contigu-ous The first criterion ensures basic load balancing whilethe second reduces capital expenditures to connect (eg viaoptics fibre) the antenna sites to the nodes As these crite-ria aim at maximizing the performance of network slicingwe argue that they correspond to a plausible deploymentstrategy We remark that the resulting node deployment isstatic and does not change during our experiments insteadthe node resources allocated to each slice may change whenemploying dynamic resource allocation schemes

Under these criteria the problem of associating the level-ℓnodes with the original antenna sites in Figure 7 is a specialcase of balanced graph k-partitioning Let us consider a graphwhere each vertex v isin V maps to one antenna site and hasan associated cost c (v ) equal to the mobile traffic demandrecorded at the site also let an edge e = uv isin E connectvertices u and v only if the corresponding antenna sites aregeographically adjacent2 The problem of level-ℓ node-to-antenna site association translates into dividing the graphinto Nℓ sub-graphs such that the sum of costs of nodes ineach partition is balanced We introduce decisions variables

euv =

1 if e is a cut edge0 otherwise

foralle isin E (5)

xvk =

1 if v is in partition k

0 otherwiseforallv isin V forallk (6)

and formulate an Integer Linear Programming (ILP) problem

minsum

euv isinE

euv (7)

stsumv isinV

xvkc (v ) le (1 + ϵ )sumv isinV c (v )

Nℓ forallk (8)

sumv isinV

xvkc (v ) ge (1 minus ϵ )sumv isinV c (v )

Nℓ forallk (9)sum

k

xvk = 1 forallv isin V (10)

euv ge xuk minus xvk foralle isin Eforallk (11)euv ge xvk minus xuk foralle isin Eforallk (12)

The objective function given by Equation (7) aims at min-imizing the number of cut edges that join vertices in sepa-rate partitions so as to generate graph subsets that are ascompact as possible Our goal in terms of load balancing isensured by the constraints given by Equations (8) and (9)which bound the load difference among the various subsetsof antennas each partition is forced to have a total cost thatis within a fraction ϵ from the ideal case of a perfectly evencostsumv isinV c (v )Nℓ The constraint given by Equation (10)

ensures that each vertex is in exactly one partition whilethose given by Equations (11) and (12) determine the valueof decision variables euv based on whether vertices u and vbelong to a same partition as defined by xuk and xvk The resulting optimization problem is NP-hard We use

a suitably configured version of the Karlsruhe Fast FlowPartitioner (KaFFPa) heuristic [29] to solve it In doing sowe allow for a plusmn10 unbalance among the load served bynodes at every level ℓ ie ϵ = 01 in Equations (8) and (9)2Multiple notions of adjacency are possible We opt for one that leveragesthe common practice of approximating antenna coverage areas via a Voronoitessellation two sites are then adjacent if they share one Voronoi cell side

Figure 8 Association of antenna sites to level-ℓ nodesin the large metropolis scenario The plots refer to ℓ =8 (16 nodes left) ℓ = 9 (8 nodes middle) and ℓ = 10 (4nodes right) Figure best viewed in colours

ℓ 1 2 3 4 5 6 7 8 9 10 11 12Traffic per node 5 10 15 30 60 75 100 150 300 600 1167 2334

NℓMetropolis 422 230 160 80 40 32 23 16 8 4 2 1City 122 60 40 20 10 8 6 4 2 1

Table 1 Hierarchical network deployments in our twourban case studies Rows are (i) the level ℓ isin 1 12(ii) the corresponding normalized mobile traffic pernode and (iii)-(iv) the number of nodes Nℓ serving areference urban region at network level ℓ At ℓ = 1nodes map to individual 4G antenna sectors and thetraffic per node is an average From ℓ = 2 to ℓ = L weconsider the partitions obtained by solving the opti-mization problem given by Equation (7)

Figure 8 shows three examples of antenna site partitioningamong network nodes for a selection of levels ℓ in the largemetropolis scenario3 Table 1 summarizes instead the mainfeatures of the partitions obtained in our two urban scenarios

4 DATA-DRIVEN EVALUATIONWe organise our evaluation as follows First we investigateworst-case settings where very stringent slice specificationsare enforced and no dynamic reconfiguration of resourcesis possible (Section 41) We then relax these constraints andassess efficiency as slice specifications are softened (Sec-tion 42) or in presence of periodic resource orchestration(Section 43) Finally we evaluate the impact of varied sliceconfigurations (Section 44) and of a resource assignmentaccounting for instantaneous traffic demands (Section 45)

41 Slicing efficiency in worst-case settingsThe least efficient sliced network scenario involves (i) strictslice specifications where the mobile network operator com-mits to guarantee the whole traffic demand (f = 1) averagedover short time periods (w = 5 minutes) for all slices and(ii) no possibility of resource reconfiguration over time ieτ spans the whole three-month observation time in our mea-surement data and |Tτ | = 1 In these worst-case settings the3Note that graph partitioning is only used to outline plausible deploymentswhere node load is reasonably balanced yet as we do not require a perfectbalance the specific partitioning algorithm is of no particular relevance

100 101 102 103

02040608

1

Mul

tiple

xing

effi

cienc

y

Large metropolisMedium-sized city=1 node

1 2 4 6 8 9 10 11 12

Figure 9 Efficiency of slice multiplexing versus thenormalized mobile traffic served by one node (bottomx axis) at level ℓ (top x axis) in the two reference ur-ban scenarios Results are for a static resource assign-ment ie |Tτ | = 1 and slice specification z = ( f w ) =(1 5minutes) Dots denote ℓ = 1 and triangles ℓ = Lfor each scenario Scattered grey points around ℓ = 1denote the efficiency and trafficmeasured at all level-1nodes (ie individual 4G antenna sectors) separately

operator is forced to replicate physical resources for differentslices statically allocating to each slice the resources neededto meet the associated offered loadThe multiplexing efficiency of slicing under these condi-

tions is presented in Figure 9 which portrays it as a functionof the network hierarchy level ℓ for the sake of clarity thelatter is also mapped to the normalized mobile traffic demandobserved by a level-ℓ node as per Table 1 Each curve refersto a urban region and confirms the intuition that the effi-ciency grows as one moves from the antenna level (dot onthe left) to a fully centralized cloud (triangle on the right)

The underlying reason for this trend is that the traffic de-mands for each slice can be very bursty at individual antennasectors this forces the allocation of substantial resourcesin order to accommodate for each slice extemporaneousactivity peaks that occur erratically in time Aggregatingdemands over an increasing number of antennas results in-stead in growingly smoother time series To substantiate thisexplanation we look into (i) the timing behaviour of thedifferent services and (ii) the impact of aggregating trafficat different levels in the network and observe the following(i) Different slices typically peak at different times eg

some during work hours and others in the evening This is ex-emplified by the time series in the left plot of Figure 4 and isin line with recent analyses of mobile service dynamics [18](ii) The burstiness of demands associated to each slice

is significantly reduced as the network level grows For in-stance in the metropolis case study the coefficients of varia-tion of the traffic time series range in [1487 2363] for ℓ = 1in [0618 0758] for ℓ = 5 and in [0511 0587] for ℓ = L

Ultimately non-aligned and elevated traffic peaks make astatic resource allocation inefficient at low network levels

For higher values of ℓ the peak intensity is reduced mitigat-ing these effects and increasing multiplexing efficiencyIn addition to the general trend of efficiency with ℓ Fig-

ure 9 allows appreciating the following quantitative results

bull The efficiency is extremely low (sim015) at the antennalevel ensuring physical resource isolation across slicesin absence of dynamic reconfiguration capabilitieswould require approximately 7 times the capacity ofa legacy architecture where no network slicing is im-plemented The grey dots in the figure highlight thatsuch poor efficiency uniformly affects all 4G antennasectors independently of their offered loadbull The efficiency grows slowly when aggregating trafficat the network edge (ℓ = 2 to ℓ = 6) Instead themultiplexing gain starts to be appreciable as onemovesabove ℓ = 7 in our reference scenarios ie at networknodes that accommodate the demands from many tensof antenna sectors at leastbull However in absolute terms even when consideringthat all traffic generated in each of our two target urbanscenarios is aggregated at a single level-L node (recallthat ℓ = L = 12 in the large metropolis and ℓ = L = 10in the medium-sized city see Table 1) the efficiencyremains fairly low at 04ndash065 In other words imple-menting the most basic form of slicing within the net-work core cloud (type-A slicing in Figure 1) would stilldouble the amount of required resources with respectto a legacy non-sliced case

Interestingly differences are minimal between the two refer-ence cities and only emerge for high values of ℓ we imputethose to the intrinsic topological and demographic differ-ences that characterize the two scenarios

The results can be disaggregated for downlink and uplinktraffic as shown in Figure 10 The outcome is consistentin the two urban regions and neatly tells apart the two di-rections Downlink traffic dominates the total demand aspreviously seen in Figure 5 therefore the associated effi-ciency curves are very close to those in Figure 9 Howeverthis is not the case for the uplink direction slicing uploadstends to become remarkably (30 to 50) less efficient as onemoves towards more centralized network levels We arguethat the reason lies again in the small uplink traffic volumewhich results in bursty time series with high peak-to-averageratios even upon aggregation over multiple antennas

The distinct trends for downlink and uplink are especiallyimportant in the light of the different costs associated to thedemands in the two directions By looking at the sheer trafficload the overall resource assignment should be driven bythe downlink behaviour since it currently dominates theaggregate data volumes as per Figure 5 However specificapplications hence slices heavily rely on uplink traffic for

002040608

1

Mul

tiple

xing

effi

cienc

y

downlinkuplink

1 2 4 6 8 9 10 11 12

002040608

1

Mul

tiple

xing

effi

cienc

y

downlinkuplink

1 2 4 6 8 9 10 11 12

100 101 102 103

02040608

1

Mul

tiple

xing

effi

cienc

y

downlinkuplink

Figure 10 Efficiency of slice multiplexing in the samesettings of Figure 9 separating downlink and uplinkTop large metropolis Bottom medium-sized city

instance the fact that efficiency at the antenna level is alsolow in uplink means that services that pose strong require-ments on access network latency (eg mobile gaming) are ashard to accommodate as the bandwidth-eager ones in down-link (eg video streaming) As another example basebandprocessing at a virtualized radio access is remarkably moreCPU-intensive for uplink traffic [8] the very low efficiencyrecorded in uplink at the network edge can make the re-sources assignment problem very challenging when dealingwith type-C type-D or type-E slices in Figure 1

42 Moderating slice specificationsThe poor efficiency found above is also caused by the verysevere slice specifications we considered To gain insighton this we investigate the impact of the QoS requirementsfor each slice on the opportunities for multiplexing slicedemands still under a static allocation of resources

We first relax the stringent requirement considered beforein the fraction f of time during which the traffic demandfor a slice must be guaranteed by the operator The left plotsin Figure 11 show how reducing f from 1 to 09 affects theefficiency of slice multiplexing at different network levels ℓand in the two reference scenarios Decreasing f drasticallyimproves the efficiency for instance by reducing the guar-anteed time percentage from 100 to slightly lower valuessuch as 995 we can nearly double the efficiency On thedownside there exists a diminishing returns effect as f islowered Even allowing an overindulgent 90 guaranteedtime percentage cannot bring efficiency above 08 for ℓ = 1the operator shall still increase its radio access capacity by20 in order to isolate slices These observations hold for allnetwork levels ℓ and in both urban regions

09 092 094 096 098 1Guaranteed time fraction f

002040608

1

Mul

tiple

xing

effi

cienc

y = 1 = 9 = L

002040608

1M

ultip

lexi

ng e

fficie

ncy = 1 = 9 = L

002040608

1

Mul

tiple

xing

effi

cienc

y = 1 = 9 = L

5 m 30 m 1 h 2 hWindow size w

002040608

1

Mul

tiple

xing

effi

cienc

y = 1 = 9 = L

002040608

1

Mul

tiple

xing

effi

cienc

y = 1 = 7 = L

002040608

1

Mul

tiple

xing

effi

cienc

y = 1 = 7 = L

Figure 11 Efficiency of slice multiplexing versus slice specifications Left guaranteed time fraction f Right aver-aging window lengthw Dashed and solid coloured lines denote the extreme network levels ℓ = 1 and ℓ = L whilethe black solid line follows an intermediate network level Top Large metropolis Bottom medium-sized city

The other parameter governing our slice specifications isthe time window lengthw over which traffic is averaged Wefind thatw has a less significant impact on efficiency than f The exact figures are in the right plots of Figure 11 the gainis mild even for long 2-hour windows and tuningw cannotreduce the large gap between the efficiency at the antennalevel and in the network core cloud Thus a 3-fold capacityincrease would be needed to implement slicing at physicallevel even ifw were set to tolerant order-of-hour valuesA final relevant aspect is that with the proposed slice

specification it is possible that the slice demands are notsatisfied over periods involving more than one consecutivetime window By appropriately setting the window size andthe f parameter we have some control over the durationof such periods For instance for the medium-sized city sce-nario and a window size ofw = 5 m the length of a periodnot fully meeting the demand is (on average) around 2 win-dows for f = 099 25 windows for f = 095 and 3 windowsfor f = 09 Similar trends are observed for the large cityand other window sizes This shows that the efficiency gainsresulting from decreasing f do not only involve a price interms of the total time not satisfying the demand but alsoin terms of the duration of the corresponding periods

43 Orchestrating resources dynamicallyWe now relax the constraint on the fully static allocation ofresources and consider a network where resources can bedynamically re-allocated to VNFs over time Such a systemallows the operator to re-assign the amount of resourcesdedicated to each slice adapting them to the actual time-varying demand for the services associated to the slice

As discussed in Section 23 we consider that the opera-tor can reconfigure the resources with a fixed periodicity τ

1 2 4 6 8 9 10 11 12

0

100

200

300

400

Perc

ent g

ain

()

100 101 102 103

02

04

06

08

1M

ultip

lexi

ng e

fficie

ncy

Large metropolisMedium-sized cityGain over static

Figure 12 Efficiency of slice multiplexing (left y axis)and percent gain over static assignment (right y axis)versus the normalized mobile traffic served by onenode (bottom x axis) at level ℓ (top x axis) in thetwo reference urban scenarios Results are for a dy-namic resource assignment where re-configurationsoccur with periodicity τ = 30 minutes under slicespecification z = ( f w ) = (1 5minutes) Dots denoteℓ = 1 and triangles ℓ = L for each scenario

which depends on the capabilities of the underlying virtu-alization technology In our scenario the operator allocatesresources optimally with respect to the target slice specifica-tions for each reconfiguration interval of duration τ This isequivalent to assuming availability of an oracle algorithmthat at the beginning of a reconfiguration interval has per-fect knowledge of the future time series of the demand foreach service and for the rest of the interval Then exactinformation about the following timespan τ allows for anoptimal matching of minimum resources to requirements asdetailed in Section 23 and exemplified in Figure 3Our baseline result in Figure 12 refers to the case of

τ = 30minutes Note that this can be regarded as a fairly high

30 m 1 h 2 h 8 h 24 h 1 w 3 MReconfiguration interval span

002040608

1

Mul

tiple

xing

effi

cienc

y = 1 = 9 = L

002040608

1M

ultip

lexi

ng e

fficie

ncy = 1 = 9 = L

002040608

1

Mul

tiple

xing

effi

cienc

y = 1 = 7 = L

Figure 13 Efficiency of slice multiplexing versus theresource reconfiguration periodicity τ Dashed andsolid coloured lines denote the extreme network lev-els ℓ = 1 and ℓ = L while the black solid line followsan intermediate network level Top Large metropolisBottom medium-sized city

resource reconfiguration frequency for several scenariosFor instance VNF management in the network core cloudhas typically larger time scales of hours or even days [32]At radio access instead faster dynamic reassignments aretechnically possible however forecasting the demand overshort time scales of minutes is challenging and easily leads toslice specification violations hence reconfiguration intervalsin the order of hours are more credible [30]

We can see from the results that dynamic allocation mech-anisms and a perfect prediction of the demand over the future30 minutes can substantially improve the efficiency of slicemultiplexing Indeed when comparing the curves in Fig-ure 12 with their equivalent in Figure 9 the gain is evidentWe made the benefit explicit as the grey region in Figure 12it ranges between 60 and 400 depending on the networklevel ℓ considered We further observe that there is a veryimportant difference between efficiency at the radio accessand in the network core A high-frequency dynamic orches-tration of resources allows for near-perfect slice multiplexingat a cloud datacenter that fully centralizes the traffic in ourlarge metropolis scenario In contrast efficiency is stuck at06 (despite a much higher percent gain) for levels close toℓ = 1 ie at individual antenna sectors or at nodes servingsmall groups of a few antennas each this implies that theoperator still has to almost double the capacity to isolateslices at network hierarchy levels close to the radio accessA more comprehensive picture is provided by Figure 13

which encompasses a wide set of reconfiguration intervalsτ from the 30 minutes case we just analysed in detail upto 3 months (ie the entire timespan of the dataset whichmaps to the static resource configuration case considered

30 m 1 h 2 h 8 h 24 h 1 w 3 M0

02040608

1

Mul

tiple

xing

effi

cienc

y = 1 = 9 = L

30 m 1 h 2 h 8 h 24 h 1 w 3 M0

02040608

1

Mul

tiple

xing

effi

cienc

y = 1 = 9 = L

002040608

1

Mul

tiple

xing

effi

cienc

y = 1 = 7 = L

Figure 14 Efficiency of slice multiplexing in pres-ence of 7 slices dedicated to specific service categoriesDashed and solid colored lines denote the extreme net-work levels ℓ = 1 and ℓ = L while the black solid linefollows an intermediate network level Top Large me-tropolis Bottom medium-sized city

in Section 41) As one could expect the multiplexing effi-ciency of slices is decreased as τ grows since the systembecomes less flexible Interestingly the loss of efficiency ismost remarkable for low values of τ reducing the frequencyof reallocation from once every 30 minutes to once every 2hours yields a high loss of efficiency (close to 02) comparableto that incurred eg by increasing τ from 2 to 8 hours If wefurther constrain the frequency of resource reallocation toonce per week or once every three months the additionalerosion of efficiency is much lower The takeaway message isthat either the operator is able to deploy virtualization tech-nologies that allow for fast reconfiguration (in the order of afew hours at most) or it is probably not worth consideringdynamic resource allocation at all

44 Varying slice configurationsThe mapping of services into specific network slice instancesmay be based on several factors such as the requirementsof the services or the specific policies implemented by eachoperator [3] The number of slices and the resulting volumeof traffic in each slice will have an impact on the overallmultiplexing efficiency which we investigate next

We first study a slice configuration where the services ofa similar type are aggregated together into the same slicewhich allows to reduce the 38 slices that we had in the pre-vious experiments down to 7 slices dedicated to streamingsocial network web cloud gaming messaging and miscella-neous services respectively Figure 14 illustrates the multi-plexing efficiency achieved by such a slice configuration as afunction of the reconfiguration period τ The values are sub-stantially larger than those obtained with a larger number of

30 m 1 h 2 h 8 h 24 h 1 w 3 M0

02040608

1

Mul

tiple

xing

effi

cienc

y = 1 = 9 = L

2 5 10 15 20 25 30 35

02040608

1

0Mul

tiple

xing

effi

cienc

y = 1 = 9 = L

2 5 10 15 20 25 30 35Number of slices

002040608

1

Mul

tiple

xing

effi

cienc

y = 1 = 7 = L

Figure 15 Efficiency of slice multiplexing as a func-tion of the number of slices x when the x minus 1 serviceswith the highest traffic load have a dedicated slice andthe remaining services are aggregated into a commonslice Dashed and solid colored lines denote the ex-treme network levels ℓ = 1 and ℓ = L while the blacksolid line follows an intermediate network level TopLarge metropolis Bottom medium-sized city

slices (see Figure 13) by aggregating service traffic we havesignificant gains in efficiency yet we lose the ability to pro-vide customized functions to each specific mobile service Itis worth highlighting however that multiplexing efficiencyremains rather low for small l and large τ valuesA second sensible slice configuration assumes that the

providers of the services that generate the highest trafficload acquire a dedicated slice tailored to their service whilethe remaining services are aggregated into a common non-customized slice In Figure 15 we analyze the multiplexingefficiency resulting from this configuration as a function ofthe total number of slices in the network (including the dedi-cated slices and the common one) when the reconfigurationperiod τ is of 1 hour and f = 1 for all slices Results showthat the trend becomes almost flat after 15 slices which im-plies that efficiency is only improved when the services withthe largest demands are brought into the common sliceIn the above slice configuration it may be reasonable to

expect that those tenants acquiring dedicated slices are pro-vided a stricter guarantees than the ones in the commonslice In order to evaluate the benefits resulting from such astrategy Figure 16 illustrates the resource savings resultingfrom providing the common slice with a guaranteed timefraction f = 09 computed as the relative percentage ofresources spared with respect to those required in the con-figuration where all slices have f = 1 Results show thatsavings remain very low in the network core (when ℓ sim L)but can be significant for resources located close to the radio

2 5 8 11 14 17 20 23 26 29 32 35 38

1020304050

0

Savi

ngs (

)

= 1 = 9 = L

2 5 8 11 14 17 20 23 26 29 32 35 38

1020304050

0

Savi

ngs (

)

= 1 = 9 = L

2 5 8 11 14 17 20 23 26 29 32 35 38Number of slices

01020304050

Savi

ngs (

)

= 1 = 7 = L

Figure 16 Savings obtained by relaxing the serviceguarantees of the common slice corresponding to thedifference between the resources required when f = 1for the common slice and those requiredwhen f = 09for that slice Dashed and solid colored lines denotethe extreme network levels ℓ = 1 and ℓ = L while theblack solid line follows an intermediate network levelTop Large metropolis Bottom medium-sized city

access (when ℓ sim 1) In the latter case savings are important(up to 20-40) when the top-10 services are included in thenon-customized low-QoS common slice Indeed as theseaccount for 65 of the overall traffic (see Figure 5) they havea much higher incidence on the system performance

45 Equipment deployment efficiencyTo conclude our analysis we look at the problem of resourcemultiplexing efficiency in a sliced network from a ratherdifferent perspective Equations (2) and (3) derived in Sec-tion 2 assume that the relevant metric for the operator isthe amount of resources utilized to accommodate the de-mand for mobile services aggregated over time Thereforethe analysis carried out in Sections 41ndash44 is appropriate toevaluate operating expenses (OPEX) which increase whenthe available resources are used more intensively and canbe applied eg to electric power consumption managementoverheads or deterioration of assets with use

However another interesting viewpoint on efficiency is interms of equipment to be deployed to meet the instantaneousdemand This relates to the capital expenditure (CAPEX) in-curred by the mobile network operator typically hardwareand infrastructure costs In this case the expressions areslightly different and capture the fact that the equipmentmust be dimensioned so as to match the peak demand For-mally let r zcs (n) be the resources needed to satisfy specifica-tions z for slice s isin S at node c isin Cℓ during reconfigurationinterval n isin T computed as indicated in Section 23 Then

002040608

1

Mul

tiple

xing

effi

cienc

y = 1 = 9 = L

002040608

1M

ultip

lexi

ng e

fficie

ncy = 1 = 9 = L

002040608

1

Mul

tiple

xing

effi

cienc

y = 1 = 7 = L

Figure 17 Efficiency of slice multiplexing from anequipment deployment perspective versus τ Dashedand solid colored lines denote the extreme networklevels ℓ = 1 and ℓ = L while the black solid line followsan intermediate network level Top Large metropolisBottom medium-sized city

the equipment resources needed to accommodate the trafficactivity peak in slice s at network level ℓ are computed as

R⋆zℓτ =sums isinS

sumc isinCℓ

maxnisinT

(r zcs (n)

) (13)

Similarly the equivalent resources needed under perfectsharing in the same settings are

P⋆zℓτ =sumc isinCℓ

maxnisinT

(r zc (n)

) (14)

where r zc (n) is the amount of resources needed to accommo-date the total demand aggregated over all slices in S at nodec and reconfiguration interval n under requirements z Themultiplexing efficiency for deployed equipment is then

E⋆zℓτ = P⋆zℓτ R

⋆zℓτ (15)

The equipment deployment efficiency given by the aboveequation is shown in Figure 17 The figure summarizes re-sults in our reference urban scenarios under a wide rangeof reconfiguration time interval durations τ and across allnetwork architectural levels ℓ We highlight the followingaspects(i) In absence of mechanisms that allow for dynamic re-

configuration the efficiency is very much comparable to thatobserved in the previous analysis as shown by the values forτ = 3 months in Figures 13 and 17 This is a clear indicationthat deploying hardware and infrastructure to provide re-source isolation across slices risks to have an unbearable costfor operators if no dynamic resource reallocation is possible(ii) Flexibility in the orchestration of resources pays off

also in terms of equipment deployment efficiency which canbe increased up to 08ndash095 when fast reconfiguration over

30-minute intervals is possible These values correspond toan additional 5ndash25 cost in terms of network infrastructureover the perfect sharing benchmark

(iii) The main difference between efficiency of resource us-age given by Equation (4) and equipment deployment givenby Equation (15) is observed at architectural levels closer toradio access When ℓ is close to 1 a dynamic reconfigurationof resources allows improving deployed infrastructure effi-ciency much faster than resource usage efficiency In otherwords resource isolation across slices has a sensibly lowerimpact on equipment installation costs than on operatingexpenses For instance at the antenna level (ℓ = 1) efficiencyis 06 in Figure 13 and 08 in Figure 17 implying that theextra cost over perfect sharing is high for resource utiliza-tion (over 60) and much lower for equipment deployment(below 25)

(iv) In contrast to the above in the network core (ie for ℓthat tends to L) trends are similar in Figure 13 and Figure 17Overall our results stress how multiplexing efficiency

of slice resources is largely consistent across the differentperspectives entailed by the expressions of Equations (4)and (15) That is the OPEX and CAPEX incurred by theoperators to support network slicing have comparable trendswith respect to the different system parameters with thenotable exception of lower deployment costs for a radioaccess infrastructure supporting high reconfigurability

5 RELATEDWORKMulti-service networks [25] are the fundamental buildingblock for the implementation of the network slicing para-digm [6] that in turn will enable new business models suchas multi-tenancy [28] and finally pave the way to 5GAt this stage the bulk of the work on next generation

network sharing architectures is already available rangingfrom novel visions of the network [24] to specific architec-tures proposals [21 35] More specifically research workalready addressed the extension to multi-service settingsof fundamental parts of the 5G system such as the RadioAccess Network (RAN) [5 12] the core network [27] orthe management and orchestration components [19] As amatter of fact that research effort is already making its wayinto standardization 3GPP is considering multi-service andnetwork slicing aspects for the next Release 15 expected todeliver the first set of 5G standards [2]

On top of the architectural research work enabling multi-service network has also been considered from an algorith-mic point of view The focal point of the research in the areahas been the resource allocation in the RAN [9 11 16 22]as the spectrum is the most difficult part of the network tooversubscribe However resource sharing in a virtualized net-work has also been tackled for other kinds of functions [14]

Despite the attention that multi-service networks net-work slicing and multi-tenant networks have been receivingfor the last few years little attention has been paid to howsuch network slices will behave in practical scenarios Un-derstanding the system efficiency in the wild has only beenpossible in reduced scenarios involving very few devices [11]or bymaking assumption on the real patterns modelling usermovements and service requests with random processes [7]

Our work sheds light on this overlooked aspect by provid-ing an empirical evaluation of slicing efficiency in large-scalescenarios in presence of realistic multi-service demands

6 TAKEAWAYS AND PERSPECTIVESWe analyzed from an empirical perspective the implicationsof real-world mobile service usage patterns on the networkinfrastructure To the best of our knowledge this is the firstattempt at understanding the impact on resource manage-ment of network slices in a multi-service multi-tenant net-work at scale We retain a number of takeaways listed nextMulti-service requires more resources Building a net-work that is capable of providing different services (possiblyassociated to several tenants) will necessarily introduce adecrease in the efficiency of the resource usage We quantifythis loss in almost one order of magnitude if consideringdistributed resources (such as spectrum) yet the efficiencyloss stays as high as 20 even in a fully centralized scenario(ie a large datacenter in the core network) These figurestranslate into high costs for the infrastructure provider whomust compensate for them by aggressively monetizing onthe new business models enabled by a multi-service scenario(eg Network Slice as a Service Infrastructure as a Service)Traffic direction is a factor Uplink and downlink trafficexhibits similar efficiency trends across network levels butuplink exacts a much higher efficiency degradation to meetequivalent QoS requirements Although uploads account fora small fraction of the overall load the further reduced effi-ciency of uplink may entail real challenges for the operatorsIndeed uplink QoS requirements are key to specific serviceswith stringent network access needs (eg mobile gaming)Even more so it is likely that multiple instances of suchservices belonging to different tenants (eg video-gamingplatforms owned by different gaming providers) have to beserved in a resource-isolated fashion in parallelLoose service level agreements may not help Althoughthe slice specifications granted to tenants may be moderatedthe overall efficiency grows only when requirements are verymuch lowered up to a point that they may be not suitablefor certain services (needing eg ldquofive nines reliabilityrdquo orbandwidth guarantees over very short time windows)Dynamic resource assignment must also be rapid Thedesign of dynamic resource allocation algorithms is crucial

to increase the efficiency of future sliced networks How-ever substantial gains will only be attained if the virtualiza-tion technologies enable a fast enough re-orchestration ofnetwork resources While current Management and Orches-tration (MampO) frameworks provide such capabilities intel-ligent algorithms able to forecast mobile service demandsand anticipate resource reconfiguration are also requiredwhich may be challenging for short timescales Underesti-mation of resources may lead to SLA violations whereasover-provisioning may harm the economic feasibility ofthe system Artificial intelligence and machine learning arepromising techniques to accomplish this [10 34] and arebeing brought into the network management landscape bystandards [10]Aggregating services is beneficial Aggregating similarservices into the same slice increases the system efficiencysignificantly yet this comes at the price of losing the ability toprovide a customized treatment to each service In contrastif the services with the highest traffic load acquire their ownslice and the remaining ones are aggregated into a commonslice the resulting gains are limited unless the common sliceincludes services with significant loadDeployment is slightly more efficient than operationWe analyzed the sharing efficiency from both a continuousresource usage and an infrastructure deployment perspectiveWhile they have similar trends in the network core theefficiency at the radio access is higher for installed hardwarein presence of high-frequency resource reallocationUrban topography has limited impact The fact that ourresults are very consistent in two urban areas of a quitedifferent nature lets us provide general insights that holdbeyond one particular scenario More precisely as usagedemands are eventually driven by human factors we expectthat our considerations may be extended to other regionsand countries in (and possibly beyond) EuropeThere is room for improvement As a final remark wewould like to stress that ours does not pretend to be a com-prehensive analysis rather one that lays the foundations toa better understanding of the new trade-offs introduced bynetwork slicing in terms of resource management efficiencyThe empirical bounds we derived represent a starting pointfor deeper investigations of a unexplored subject with strongimplications for the future generations of mobile networks

ACKNOWLEDGMENTSWe would like to thank the shepherd and reviewers fortheir valuable comments and feedback The work of Uni-versity Carlos III of Madrid was supported by the H20205G-MoNArch project (grant agreement no 761445) and thework of NEC Laboratories Europe was supported by theH2020 5G-Transformer project (grant agreement no 761536)

REFERENCES[1] 3rd Generation Partnership Project (3GPP) 2015 Cellular system sup-

port for ultra-low complexity and low throughput Internet of Things(CIoT) 3GPP Technical Report (TR) 45820

[2] 3rd Generation Partnership Project (3GPP) 2018 NR and NG-RANOverall Description Stage-2 (Release 15) 3GPP Technical Specification(TS) 38300

[3] 3rd Generation Partnership Project (3GPP) 2018 Telecommunicationmanagement Study on management and orchestration of networkslicing for next generation network (Release 15) 3GPP TechnicalReport (TR) 28801

[4] 5th Generation Public Private Partnership (5G-PPP) 2017 View on5G Architecture (version 20) 5G-PPP Architecture Working GroupWhite Paper

[5] I F Akyildiz P Wang and S Lin 2015 SoftAir A software definednetworking architecture for 5G wireless systems Computer Networks85 (July 2015) 1ndash18

[6] Next GenerationMobile Networks (NGMN) Alliance 2015 Descriptionof network slicing concept NGMN White Paper

[7] D Bega M Gramaglia A Banchs V Sciancalepore K Samdanis and XCosta-Perez 2017 Optimising 5G infrastructure markets The businessof network slicing In Proceedings of the IEEE International Conferenceon Computer Communications (IEEE INFOCOM 2017) Atlanta GA

[8] S Bhaumik S P Chandrabose M K Jataprolu G Kumar A Muralid-har P Polakos V Srinivasan and T Woo 2012 CloudIQ a frameworkfor processing base stations in a data center In Proceedings of the 18thAnnual International Conference on Mobile Computing and Networking(ACM MobiCom 2012) Istanbul Turkey

[9] P Caballero A Banchs G de Veciana and X Costa-Perez 2017 Net-work slicing games Enabling customization in multi-tenant networksIn Proceedings of the IEEE International Conference on Computer Com-munications (IEEE INFOCOM 2017) Atlanta GA

[10] European Telecommunications Standards Institute (ETSI) 2017 Im-proved operator experience through Experiential Networked Intelli-gence (ENI) Introduction - Benefits - Enablers - Challenges - Call forAction ETSI White Paper No 22

[11] X Foukas M K Marina and K Kontovasilis 2017 Orion RANSlicing for a Flexible and Cost-Effective Multi-Service Mobile NetworkArchitecture In Proceedings of the 23rd Annual International Conferenceon Mobile Computing and Networking (ACM MobiCom 2017) SnowbirdUT

[12] X Foukas N Nikaein M M Kassem M K Marina and K Kontovasilis2016 FlexRAN A Flexible and Programmable Platform for Software-Defined Radio Access Networks In Proceedings of the 12th Internationalon Conference on Emerging Networking EXperiments and Technologies(ACM CoNEXT 2016) Irvine CA

[13] Google [n d] Google Project Fi httpsfigooglecomabout[14] J G Herrera and J F Botero 2016 Resource Allocation in NFV A

Comprehensive Survey IEEE Transactions on Network and ServiceManagement 13 3 (Sept 2016) 518ndash532

[15] A Ksentini and N Nikaein 2017 Toward Enforcing Network Slicingon RAN Flexibility and Resources Abstraction IEEE CommunicationsMagazine 55 6 (June 2017) 102ndash108

[16] Y L Lee J Loo T C Chuah and L C Wang 2018 Dynamic NetworkSlicing for Multitenant Heterogeneous Cloud Radio Access NetworksIEEE Transactions on Wireless Communications 17 4 (April 2018) 2146ndash2161

[17] X Li D Li J Wan A V Vasilakos C Lai and S Wang 2017 A reviewof industrial wireless networks in the context of Industry 40 WirelessNetworks 23 1 (Jan 2017) 23ndash41

[18] C Marquez M Gramaglia M Fiore A Banchs C Ziemlicki and ZSmoreda 2017 Not All Apps Are Created Equal Analysis of Spa-tiotemporal Heterogeneity in Nationwide Mobile Service Usage InProceedings of the 13th International Conference on Emerging Network-ing EXperiments and Technologies (ACM CoNEXT 2017) IncheonSeoulSouth Korea

[19] A Mayoral R Vilalta R Casellas R Martinez and R Munoz 2016Multi-tenant 5G Network Slicing Architecture with Dynamic Deploy-ment of Virtualized Tenant Management and Orchestration (MANO)Instances In Proceedings of the 42nd European Conference and Exhibi-tion on Optical Communication (ECOC 2016) Dusseldorf Germany

[20] T L Nguyen and A Lebre 2017 Virtual Machine Boot Time Model InProceedings of the 25th Euromicro International Conference on ParallelDistributed and Network-based Processing (PDP 2017) St PetersburgRussia

[21] N Nikaein E Schiller R Favraud K Katsalis D Stavropoulos IAlyafawi Z Zhao T Braun and T Korakis 2015 Network StoreExploring Slicing in Future 5G Networks In Proceedings of the 10thInternational Workshop on Mobility in the Evolving Internet Architecture(ACM MobiArch 2015) Paris France

[22] B Niu Y Zhou H Shah-Mansouri and VW SWong 2016 ADynamicResource Sharing Mechanism for Cloud Radio Access Networks IEEETransactions on Wireless Communications 15 12 (Dec 2016) 8325ndash8338

[23] M Odini 2016 OpenSource MANO IEEE Softwarization A Collectionof Short Technical Articles httpssdnieeeorgnewsletterjuly-2016opensource-mano

[24] P Rost A Banchs I Berberana M Breitbach M Doll H Droste CMannweiler M A Puente K Samdanis and B Sayadi 2016 Mobilenetwork architecture evolution toward 5G IEEE CommunicationsMagazine 54 5 (May 2016) 84ndash91

[25] P Rost C Mannweiler D S Michalopoulos C Sartori V Sciancale-pore N Sastry O Holland S Tayade B Han D Bega D Aziz andH Bakker 2017 Network Slicing to Enable Scalability and Flexibilityin 5G Mobile Networks IEEE Communications Magazine 55 5 (May2017) 72ndash79

[26] O Sallent J Perez-Romero R Ferrus and R Agusti 2017 On Radio Ac-cess Network Slicing from a Radio Resource Management PerspectiveIEEE Wireless Communications 24 5 (Oct 2017) 166ndash174

[27] M R Sama X An Q Wei and S Beker 2016 Reshaping the mobilecore network via function decomposition and network slicing for the5G Era In Proceedings of the 2016 IEEE Wireless Communications andNetworking Conference (IEEE WCNC 2016) Doha Qatar

[28] K Samdanis X Costa-Perez and V Sciancalepore 2016 From net-work sharing to multi-tenancy The 5G network slice broker IEEECommunications Magazine 54 7 (July 2016) 32ndash39

[29] P Sanders and C Schulz 2013 Think Locally Act Globally HighlyBalanced Graph Partitioning In Proceedings of the International Sym-posium Experimental Algorithms (SEA 2013) Rome Italy

[30] V Sciancalepore K Samdanis X Costa-Perez D Bega M Gramagliaand A Banchs 2017 Mobile traffic forecasting for maximizing 5Gnetwork slicing resource utilization In Proceedings of the IEEE Interna-tional Conference on Computer Communications (IEEE INFOCOM 2017)Atlanta GA

[31] S K Sharma T E Bogale L B Le S Chatzinotas X Wang and BOttersten 2018 Dynamic Spectrum Sharing in 5G Wireless NetworksWith Full-Duplex Technology Recent Advances and Research Chal-lenges IEEE Communications Surveys amp Tutorials 20 1 (Feb 2018)674ndash707

[32] F Z Yousaf and T Taleb 2016 Fine-grained resource-aware virtualnetwork function management for 5G carrier cloud IEEE Network 302 (March 2016) 110ndash115

[33] Y Zaki T Weerawardane C Gorg and A Timm-Giel 2011 Multi-QoS-Aware Fair Scheduling for LTE In Proceedings of the IEEE 73rdVehicular Technology Conference (IEEE VTC 2011 Spring) BudapestHungary

[34] C Zhang P Patras and H Haddadi 2018 Deep Learning in Mobileand Wireless Networking A Survey (March 2018) arXiv180304311[csNI]

[35] H Zhang N Liu X Chu K Long A H Aghvami and V C M Leung2017 Network Slicing Based 5G and Future Mobile Networks Mo-bility Resource Management and Challenges IEEE CommunicationsMagazine 55 8 (Aug 2017) 138ndash145

Abstract
1 Introduction
2 Network scenario and metrics
- 21 Network slicing scenario
- 22 Slice specifications
- 23 Resource allocation to slices
- 24 Multiplexing efficiency
- - 3 Case studies
  - - 31 Mobile service demands
    - 32 Hierarchical network structure
    - - 4 Data-driven evaluation
      - 41 Slicing efficiency in worst-case settings
        
        42 Moderating slice specifications
        
        43 Orchestrating resources dynamically
        
        44 Varying slice configurations
        
        45 Equipment deployment efficiency
        
        5 Related work
        
        6 Takeaways and perspectives
        
        Acknowledgments
        
        References

Page 2: How Should I Slice My Network? A Multi-Service Empirical ... · radio access, through, e.g., cloud RAN (C-RAN) paradigms. Here, basic radio-access slices allow for tailored MAC-layer