PlaceRAN: Optimal Placement of Virtualized Network Functions in the Next-generation Radio Access Networks
Fernando Zanferrari Morais, Gabriel Matheus de Almeida, Leizer Pinto, Kleber Vieira Cardoso, Luis M. Contreras, Rodrigo da Rosa Righi, Cristiano Bonato Both
11 PlaceRAN: Optimal Placement of VirtualizedNetwork Functions in the Next-generation RadioAccess Networks
Fernando Zanferrari Morais, Gabriel Matheus F. de Almeida, Leizer Pinto,Kleber Vieira Cardoso, Luis M. Contreras, Rodrigo da Rosa Righi, and Cristiano Bonato Both (cid:70)
Abstract —The fifth-generation mobile evolution introduces Next-Generation Radio Access Networks (NG-RAN) based on the protocolstack disaggregation to enable flexibility to support users’ demand forultra-low latency and high-bandwidth applications. For example, OpenRAN solutions focus on an NG-RAN with general-purpose vendor-neutral hardware, software-defined technologies, and interoperability,splitting the protocol stack into the eight options combined into threenetwork units, i.e., central, Distributed, and Radio. These network unitsand the protocols disaggregated are managed as radio functions. Thesefunctions’ placement is challenging since the best decision is based onthe RAN protocol stack split, routing paths of transport networks withrestricted bandwidth and latency requirements, different topologies andlink capabilities, asymmetric computational resources, etc. Therefore,this article proposes the first exact model for the placement optimizationof radio functions for vNG-RAN planning, named PlaceRAN. The mainobjective is to minimize the computing resources and maximize the ag-gregation of radio functions. The PlaceRAN evaluation considered tworeal network topologies based on 5G-crosshaul and Passion Europeanprojects. The PlaceRAN performance reaches up to 80% aggregation ofradio functions centralized.
Index Terms —Placement optimization, vNG-RAN, Disaggregation
NTRODUCTION T HE fifth-generation mobile evolution is based on stan-dards [1]–[3] defining the virtualization of the RadioAccess Network (vRAN) on the protocol stack disaggrega-tion and the software-based networking. These standardsspecify the Next Generation RAN (NG-RAN) architectureto meet the new network demands, e.g., ultra-low latencyand high-bandwidth applications. Furthermore, the indus-try creates its initiatives for NG-RAN. The most promis-ing are based on open solutions, called Open RAN (O-RAN), focused on general-purpose vendor-neutral hard-ware, software-defined technologies, and interoperability[4]. The NG-RAN architecture proposed enables the basestation for splitting the protocol stack into the eight options • Fernando Zanferrari Morais, Rodrigo da Rosa Righi, and Cristiano BonatoBoth are with the University of Vale do Rio dos Sinos (UNISINOS). • Gabriel Matheus, Leizer Pinto, Kleber Vieira Cardoso are with the Uni-versidade Federal de Goi´as (UFG). • Luis M. Contreras is with the Transport & IP Networks - Systems andNetwork Global Direction, Telef´onica GCTIO Unit. combined into three network elements: (i) Central Unit(CU), (ii) Distributed Unit (DU), and (iii) Radio Unit (RU)[2], [5]. These split options concern improving the radiofunctions and cost efficiency compared with the last mobilegenerations [5], [6].The network software process in NG-RAN is guided bythe virtualization of nodes and radio functions by the Net-work Function Virtualization (NFV) concept [7], [8]. For ex-ample, the NG-RAN architecture functional split combinedwith vRAN provides flexibility for mobile access networks.This flexibility allows mobile network operators to placethe radio functions to take into account available networkresources and user demand. The placement of radio func-tions in a fine-grained network management approach isconsidered vital for fifth-generation networks to achievethe expected leadership of digital transformation. However,the development of the virtualized NG-RAN architecture(vNG-RAN) is an unprecedented challenging problem sincecrosshaul transport networks (backhaul, midhaul, and fron-thaul) have restricted bandwidth and latency requirements,different topologies and link capabilities, asymmetric com-puting resources (CR), and unbalanced user demand [2], [9].vNG-RAN is up-to-date in fifth-generation research, al-though no decision-making method for specification design-ing of the placement of virtualized radio network functionsis defined. The placement is defined on the optimizationproblem [10], [11] based on the best joint decision betweenthe split of the RAN protocol stack, the routing paths of thecrosshaul network, and the CRs strategies of the CU, DU,and RU nodes [11]. Therefore, the ideal placement leads toanalysis related to the bandwidth and latency requirementsfor each split option between CU-DU and DU-RU. More-over, each split option results in a computing cost (process-ing, memory, and storage) to be evaluated. According to thenetwork’s scalability, the placement needs to be aware ofthe load occupation, even routing and computing resourcesallocation.In the literature, several works lead to the placement op-timization of radio functions. The main strategies developedare to maximize the number of VNFs running in a single CU,DUs fixes, and close to RUs [12], [13]. Moreover, CU is co-located with the core [14]. The state-of-the-art is restricted a r X i v : . [ c s . N I] M a r considering the protocol disaggregations number, reachingthe maximum of five [15], efficiency under crosshaul con-straints (mainly, the fronthaul network) [12], [16], [17], andcomputing resources [18], [19]. Therefore, to the best of ourknowledge, no work in the literature considers CUs, DUs,and RUs for real-world networks, making the problem moregeneral with high functional split options and protocol stackanalysis. Contributions . In this article, we introduce PlaceRAN,the problem formulation for the optimal placement of vNG-RAN functions. The problem is formulated as the best trade-off between maximizing the aggregation level of virtual-ized NG-RAN functions and minimizing the number ofcomputing resources necessary for running these functions.PlaceRAN innovates by considering in the formulation allRAN elements (CU, DU, and RU), the paths between theseelements (fronthaul, midhaul, and backhaul), and also allfunctional splits according to the standards. Our contribu-tions can be summarized as follows: • New problem formulation – in some aspects, Plac-eRAN is the most general problem formulation inthe context of vNG-RAN, and it was designed witha comprehensive set of real-world NG-RANs consid-erations in mind. • New approach – we introduced some concepts toproperly formulate PlaceRAN, such as Disaggre-gated RAN Combination (DRC) and multi-stageproblem formulation, turning the problem formula-tion simple despite its generality. • Efficient exact solution – we solve PlaceRAN using aconventional solver (i.e., IBM CPLEX) for real-worldRAN instances despite the problem complexity. • Relevant results and new insights – our evaluationemployed examples of present and future RANs. Weshow how PlaceRAN can contribute to the virtualiza-tion of present RANs, but it also ready to deal withthe optimal placement of forthcoming vNG-RANs.
Article organization.
Section 2 introduces the vNG-RANbackground. The PlaceRAN system model and problemstatement are described in Section 3. Section 4 presents thePlaceRAN evaluation methodology and results. The relatedwork is discussed in Section 5, and finally, Section 6 presentsthe final remarks.
IRTUALIZED
NG-RAN P
LACEMENT
The fundamental idea of a disaggregated NG-RAN is todecompose the RAN functions into virtualized componentsthat can be distributed to run into different computingdevices, i.e., a non-monolithic approach. Therefore, it isnecessary to identify how this decomposition can be per-formed and which conditions must be satisfied to have thedisaggregated version running correctly. This disaggregatedNG-RAN is defined by the concept of functional splits thatspecifies all the network functions, clear interface points be-tween them, and the requirements of each network function[4], [20].The number of functional splits and where they occur aredetermined by specifications from standardization bodies,such as Release 14 from 3GPP [21] and IMT-2020/5G from
TABLE 13GPP Latency and bitrate requirements for each split.
Split Functional One-way Bitrate (Gbps)Option Split latency DL ULO1 RRC-PDCP 10 ms 4 3O2 PDCP - High RLC 10 ms 4 3O3 High RLC - Low RLC 10 ms 4 3O4 Low RLC - High MAC 1 ms 4 3O5 High MAC - Low MAC < µ s 4.13 5.64O7 High PHY - Low PHY 250 µ s 86.1 ∗ ∗ O8 Low PHY - RF 250 µ s 157.3 157.3O7 split maximum value. ∗ ITU-T [2]. Table 1 shows the specifications of the disaggre-gated protocol stack employed in this work. The (maximum)latency and (minimum) bitrate values must be satisfied withthe communication of the RAN nodes (CU, DU, and RU)even if they are running on different computing devices.The latency and bitrate must be assured according to thefunctional split specified. The (maximum) latency and (min-imum) bitrate values presented in the table correspond to anRU with the following configuration: 100 MHz bandwidth,32 antenna ports, 8 MIMO layers, and 256 QAM modulation[21].Each RAN node (CU, DU, and RU) is considered anetwork function virtualized as part of a disaggregatedNG-RAN (3 independents nodes) composed of RAN pro-tocols embedded in Virtualized Network Functions (VNFs).Moreover, each RAN node may be identified by the set ofall RAN protocols running into it, as shown in Fig. 1. Aconfiguration with less than three nodes may be named asDU and RU integration (DU and RU), C-RAN (CU and DU),or D-RAN (CU, DU, and RU) [2]. In a disaggregated NG-RAN, the paths along the network connecting the core tovCU, vCU to vDU, and vDU to RU are defined as backhaul,midhaul, and fronthaul, respectively. This terminology isuseful since each physical link of the access network acts asa crosshaul, meaning that it can transport any combinationof the previous paths. The crosshaul needs to ensure thelatency and bitrate required according to the functionalsplits [9], [20].As illustrated in Fig. 1, each functional split and thecorresponding placement of VNFs in a specific RAN nodecharacterize a
Disaggregated RAN Combination (DRC). Theconcept of DRC, introduced in this article, represents thepreservation of the protocol stack order during the process-ing of the VNFs. Nineteen DRCs are mapped consideringseven split options. The O8 option is not virtualized sincethe RF protocol is hardware-based, making virtualizationimpractical. Certain DRCs are not used in practice becausethey are not cost-effective, e.g., midhaul with less than 1 ms,or they lack advantages in the RAN disaggregation.We highlighted the nine DRCs effectively adopted invRAN deployments, whose choice is based on standard-ization bodies and industry alliances [1]–[3]. In both archi-tectures of three independent nodes (O-RAN and SmallcellForum), the focus is on O1 and O2 (Fig. 1 - DRC1, DRC2,DRC7, and DRC8). The industry considers split O1 as apossibility of the decentralized data plane. Split O2 is con-solidated by 3GPP and ITU-T via the F1 interface and is an
DRC1 CU + DU +RUCU
RCPDHRLRHMLMHPLPRF RCPDHRLRHMLMHPLPRF RCPDHRLRHMLMHPLPRF RCPDHRLRHMLMHPLPRF RCPDHRLRHMLMHPLPRF RCPDHRLRHMLMHPLPRFO1 O2 O3 O4 O5 O6O7 RCPDHRLRHMLMHPLPRF RCPDHRLRHMLMHPLPRF RCPDHRLRHMLMHPLPRF RCPDHRLRHMLMHPLPRF RCPDHRLRHMLMHPLPRFO1 O2 O3 O4 O5 RCPDHRLRHMLMHPLPRFO6 RCPDHRLRHMLMHPLPRFO7 RCPDHRLRHMLMHPLPRF
CURUDU
RCPDHRLRHMLMHPLPRF RCPDHRLRHMLMHPLPRF RCPDHRLRHMLMHPLPRF RCPDHRLRHMLMHPLPRF RCPDHRLRHMLMHPLPRFO1 O2 O3 O4 O5O6
CURUDU DU +RU CU +DURU
DRC2
DRC3 DRC4 DRC5 DRC6
DRC7 DRC8
DRC9 DRC10DRC11
DRC12 DRC13
DRC14 DRC15 DRC16
DRC17 DRC18 DRC19 O DRC
LR - Low RLCHM - High MACLM - Low MAC HP - High PHYLP - Low MACRF - Radio Frequency
Fig. 1. Functional split and computing devices for a disaggregated NG-RAN. industry reference for O-RAN and Smallcell Forum initia-tives [2], [22], [23]. Two industry DRCs were chosen for theDU and RU integration (Fig. 1 - DRC12 and DRC13), besidesof the two C-RAN options (Fig. 1 - DRC17 and DRC18).These splits align with ITU-T (mainly due to crosshaulconstraints) and O-RAN and Smallcell Forum initiatives [4],[22]. Naturally, the traditional D-RAN architecture is alsosupported to provide scenarios where the crosshaul is verylimited [2] (Fig. 1 - DRC19).In summary, the disaggregated NG-RAN can be im-plemented as a virtual network service, i.e., a collectionof VNFs with a particular set of characteristics. First, theservice consists of process one protocol stack per RF devicein NG-RAN. This processing implies an appropriate orderof the flow-through VNFs, i.e., Service Function Chain (SFC)to be respected. VNFs are instantiated in the RAN nodes,which are also virtual elements that can run in differentcomputing devices in NG-RAN. The choice of where toposition the RAN nodes and their VNFs affects the re-sources applied, including computing and networking. Foreach NG-RAN topology and set of resources, there may bemultiple options for positioning VNFs and RAN nodes. Ingeneral, the objective is to consume the minimum resourcesand group the maximum of VNFs related to the sameprotocol or layer. However, each positioning option impliesdifferent computing and networking demands, which mustnot exceed the available resources. Therefore, the function’splacement becomes a complex optimization problem thatwe will formally present in the next section.
ODEL AND PROBLEM STATEMENT
Initially, Subsection 3.1 presents the system model of avirtualized formally and disaggregated NG-RAN, in whichdifferent functional splits are possible. Moreover, the place-ment of the virtual functions also is introduced with multi-ple options. Second, Subsection 3.2 formulates the optimiza-tion problem of jointly minimizing the number of necessary computing resources, selecting the functional splits, andplacing the virtual functions.
Nodes and links.
According to the 3GPP standards (Release15 [1] and Release 16 [24]), we consider RAN of a mobilenetwork connection to the core network, as illustrated byFig. 2. RAN is composed of: • A set B = { b , b , ..., b |B| } of RUs, i.e., nodes hostingthe Low PHY sublayer and the RF processing basedon a lower layer functional split. • A set C = { c , c , ..., c |C| } of CRs that may processthe virtual functions. Each CR c m has a process-ing capacity c P rocm (given in some reference cores).Moreover, each CR has other characteristics, suchas memory and storage capacity, but they are notcommonly exhausted before the processing capacityin the context of disaggregated RAN. A CR mayconnect directly to an RU. • A set T = { t , t , ..., t |T | } of transport nodes, whichmay connect to RUs, CRs, core, or each other.To represent RAN and core, we define the graph G =( V , E ) , with V = { v } ∪ B ∪ C ∪ T being the set of nodesand E = { e ij ; v i , v j ∈ V } , v i , v j ∈ V} representing theset of network links connecting the nodes. v represents thecore and it is the source/destination for all flows. Each link e ij ∈ E has a transmitting capacity e Capij (given in multiplesof bps) and a latency e Latij (given in fractions of a second).
Paths and routing.
We consider that all network traffic hasthe core as its source (downlink) or destination (uplink).However, without loss of generality, we represent only thedownlink case in this work. We define P l as the set of k -shortest paths from the core to each RU b l ∈ B . Each path p ∈ P l is composed of three sub-paths: p BH (backhaul), p MH (midhaul), and p F H (fronthaul), in which at least oneof these sub-paths is not empty.
TransportnodeCRRU Core
Fig. 2. RAN considered as a reference to the system model.
Virtualized RAN functions.
We consider that a VNF runsa protocol of RAN stack (except the RF protocol, as detailedpreviously). Moreover, VNFs are labeled in increasing order,starting from PHY Low with f and ending at RRC with f . We define F = { f , f , f , f , f , f , f , f } as the setof disaggregated RAN VNFs, where the distribution mustfollow one of the industry DRCs [2], [23] of the set D = { D , D , ..., D |D| } (illustrated in Fig. 1). Our problem has two objectives: (i) maximize the aggre-gation level of RAN VNFs and (ii) minimize the numberof CRs used for this aggregation. Since the computing andnetwork capacities are limited, decreasing the number ofCRs may not imply an increase in the aggregation level,which creates conflicting objectives. However, there is aclear relationship between the number of CRs and the aggre-gation level. Additionally, the functional splits’ aggregationlevel is not measured only by the number of VNFs andCRs. The aggregation level is also affected by two othermetrics: the number of DRCs employed and the priorityor preference of each DRC. Since three incompatible metricsmeasure the aggregation level, we designed our formulationinto three stages. The optimal solution can be eventuallyobtained in the first or second stage, but we are sure aboutit only in the third stage, after resolving all potential draws.
First Stage
In the first stage, the objective is jointly to maximize thenumber of grouped RAN VNFs and minimize the numberof CRs used to run these VNFs. We define x p,rl ∈ { , } asthe decision variable representing which path p ∈ P l andDRC D r ∈ D is selected to serve RU b l ∈ B . From the inputdata, we determine u pm ∈ { , } to indicate if c m ∈ C ispart of the p ∈ P l . Additionally, we define the mappingfunction M ( c m , f s , b l , D r ) ∈ { , } over the input data,which indicates if the CR c m ∈ C runs the VNF f s ∈ F fromthe RU b l ∈ B , according to the DRC D r ∈ D . Therefore, wedefine the following objective function: minimize Φ − Φ , (1)where Φ represents the amount of CRs, given by: Φ = (cid:88) c m ∈ C (cid:38) (cid:80) b l ∈ B (cid:80) D r ∈D (cid:80) p ∈P l ( x p,rl · u pm ) |C| (cid:39) , (2)and Φ represents the amount of grouped RAN VNFs,given by: Φ = (cid:88) c m ∈C (cid:88) f s ∈F (cid:88) b l ∈B (cid:88) D r ∈D (cid:88) p ∈P l [ x p,rl · u pm · M ( c m , f s , b l , D r )] − (cid:80) D r ∈D (cid:80) p ∈P l (cid:80) b l ∈B [ x p,rl · u pm · M ( c m , f s , b l , D r )] |F| . (3)For each RU b l ∈ B , exactly one DRC D r ∈ D , using asingle path p ∈ P l , must be selected, as represented by thefollowing constraint: (cid:88) D r ∈D (cid:88) p ∈P l x p,rl = 1 , ∀ b l ∈ B . (4)The transmitting capacity e Capij of every link e ij must notbe exceeded, as described by the following constraint: (cid:88) b l ∈B (cid:88) D r ∈D (cid:88) p ∈P l (cid:104) x p,rl (cid:16) y p BH e ij · α rBH + y p MH e ij · α rMH + y p FH e ij · α rF H (cid:17)(cid:105) ≤ e Capij , ∀ e ij ∈ E , (5)where y p BH e ij , y p MH e ij , and y p FH e ij indicate if the link e ij ispart of the backhaul, midhaul, or fronthaul, respectively, ina path p ∈ P l that transports a specific DRC D r ∈ D . Each D r ∈ D has associated demands for bitrate in the backhaul,midhaul, and fronthaul, represented by α rBH , α rMH , and α rF H , respectively. There are functional splits in which thepath p ∈ P l has less than three sub-paths, e.g., DRCs 12, 17,and 19 (as illustrated in Fig. 1). Moreover, if the sub-path isabsent, then no link is part of it.Each D r ∈ D tolerates a maximum latency in each sub-path (backhaul, midhaul, and fronthaul) of the path p ∈ P l ,which is described by the following constraints: (cid:88) e ij ∈E x p,rl · y p BH e ij · e Latij ≤ β rBH , ∀ b l ∈ B , p ∈ P l , D r ∈ D , (6) (cid:88) e ij ∈E x p,rl · y p MH e ij · e Latij ≤ β rMH , ∀ b l ∈ B , p ∈ P l , D r ∈ D , (7) (cid:88) e ij ∈E x p,rl · y p FH e ij · e Latij ≤ β rF H , ∀ b l ∈ B , p ∈ P l , D r ∈ D , (8)where β rBH , β rMH , and β rF H represent the maximumlatency tolerated in the backhaul, midhaul, and fronthaul, respectively, of a path p ∈ P l that transports a specific DRC D r ∈ D . There are functional splits in which the path p ∈ P l has less than three sub-paths. In the same way, if the sub-path is absent, then no link is part of it.Finally, the VNFs selected to run in a CR c m ∈ C mustnot exceed its processing capacity c P rocm , as represented bythe following constraint: (cid:88) f s ∈F (cid:88) b l ∈B (cid:88) D r ∈D (cid:88) p ∈P l x p,rl · u pm · M ( c m , f s , b l , D r ) · γ sm ≤ c P rocm , ∀ c m ∈ C , (9)where γ sm is the computing demand of the VNF f s ∈ F . Second Stage
After solving the first stage, we obtain the minimum num-ber of CRs necessary to achieve the maximum aggregationlevel of RAN VNFs. Since these two objectives may beconflicting, the first stage’s final result is the best trade-off between these goals. However, the aggregation levelachieved may not be optimal. An example is illustrated inFig. 3, where solution 1A and solution 1B are two possiblesolvers’ outcomes. In this case, both solutions have the samevalue of object function, i.e., both achieve the same value forEquation (1). f f f f f f f f f f f f f f f f f f f f f f f f f f f Solution 1A f f f f f f f f f f f f f f f f f f f f f f f f f f f Solution 1B
TransportnodeCR Idle resourceBusy resource (VNF)RU Core Core
Fig. 3. A possible tie after solving the first stage.
While comparing different solutions with the same valuefor the objective function (Equation (1)) in the first stage,we observed that those with a smaller number of DRCscombine, sharing the same virtualized RAN function F and CR, improving the system performance. For example,in Fig. 3, there is a benefit in sharing CR 6 by f (PDCP)from the red flow and f from the green flow (solution 1A).However, there is no benefit in sharing CR 6 by f (PDCP)from the red flow and f from the blue flow (solution 1B).Solution 1A has two DRCs: 2 and 12, while solution 1B hasthree DRCs: 1, 2, and 13. Therefore, the objective function ofthe second stage is to minimize the number of DRCs: minimize (cid:88) D r ∈D (cid:38) (cid:80) b l ∈B (cid:80) p ∈P l x p,rl |B| (cid:39) . (10) This second stage must consider only solutions withexactly the same value of the objective function achievedby the optimal solution from the first stage. The followingconstraint assures this situation: Φ − Φ = f st stage ( x p,r ∗ l ) , (11)where f st stage ( x p,r ∗ l ) represents the objective func-tion’s value from the first stage when the optimal solutionis found. Additionally, all constraints from the first stagemust also be satisfied, i.e., the second stage is subject to theconstraints (4)–(9). Third Stage
After solving the second stage, we eliminate potential solu-tions that have incompatible layers sharing common CRs.While this improves the aggregation level, there is still thepossibility of obtaining different solutions with the samenumber of DRCs, but not equivalent. An example is illus-trated in Fig. 4, where solution 2A and solution 2B are twopossible solvers’ outcomes. These solutions have the samevalue of object function, i.e., both achieve the same value forEquation (10). In this case, solution 2A has two DRCs: 2 and12, while solution 2B also has two DRCs: 1 and 12. f f f f f f f f f f f f f f f f f f f f f f f f f f f Solution 1A f f f f f f f f f f f f f f f f f f f f f f f f f f f Solution 1B
TransportnodeCR Idle resourceBusy resource (VNF)RU Core Core
Fig. 4. A possible tie after solving the second stage.
The solutions in Fig. 4 are not equivalent because eachDRC has characteristics that make it unique, i.e., it is pos-sible to rank DRCs and the solution’s quality. By rankingDRCs, we are also able to differentiate the solutions again.This rank can be directly extracted from the standards,which already specify the preference order of DRCs. Nat-urally, different standardization bodies (e.g., O-RAN andITU) may assign other priorities to DRCs, but our model isgeneric and works properly with any of them. In the priorityassignment for each DRC, the smaller the value, the higherpriority. Therefore, the objective function of the third stageis to minimize the sum of values assigned to DRCs: minimize (cid:88) b l ∈B (cid:88) D r ∈D (cid:88) p ∈P l ( x p,rl · D ωr ) , (12) where D ωr represents the priority of the DRC D r .The third stage must also consider only solutions withexactly the same value of the objective function achieved bythe optimal solution from the first stage, i.e., the constraintdescribed by Equation 11 and the exact value of the objectivefunction given by the second stage’s optimal solution. Thefollowing constraint assures this definition: (cid:88) D r ∈D (cid:38) (cid:80) b l ∈B (cid:80) p ∈P l x p,rl |B| (cid:39) = f nd stage ( x p,r ∗ l ) , (13)where f nd stage ( x p,r ∗ l ) represents the value of the objec-tive function from the second stage when the optimal solu-tion is found. Similarly to the second stage, all constraintsfrom the first stage must also be satisfied, i.e., the third stageis subject to the constraints (4)–(9). VALUATION
This section evaluates the PlaceRAN model in several RANconfigurations, including T1 present and T2 future networktopologies, different amounts of resources, and distinct de-mands. Subsection 4.1 provides a general description ofthe evaluated scenarios, detailing the topologies, resources,and which parameters are varied to assess the solutionresults. Subsection 4.2 presents and discusses the resultsobtained in the evaluation of PlaceRAN, which involvesDRC selection, minimization of CRs versus maximizationaggregations level, general aspects of the solutions, andcharacteristics of the optimization model.
Table 2 summarizes the scenarios employed in the evalua-tion of the PlaceRAN model. There are three types of scenar-ios: Low Capacity (LC), Random Capacity (RC), and HighCapacity (HC). Each scenario has four types of TransportNodes, which are defined according to the proximity tothe Core and the number of neighbors: aggregation node1 (AG1), aggregation node 2 (AG2), access node 1 (AC1),and access node 2 (AC2). Based on each transport node,there are four characteristics of real networks: (i) number ofCRs, (ii) Bandwidth, (iii) Latency, and (iv) number of RUnodes. Moreover, our evaluation is focused on comparingthe scenarios and two topologies. T1 represents the currentRANs and based on the 5G-crosshaul project (in Table 2 as † ). T2 shows a trend in the design of the future RANs andaligned with the PASSION project (in Table 2 as ‡ ). Transport Nodes.
We associate the nodes according toT1 and T2 topologies. T1 provides a current operationalnetwork of 51 nodes in the shape of a ring, formed byan aggregation ring and other access rings. Whereas T2defines a trend in future RAN design with a hierarchi-cal tree structure, presenting two aggregation stages andother access stages. As suggested by the PASSION project,we limit T2 to 128 nodes. Fig. 5 illustrates the two RANtopologies employed in this evaluation. In both topologies,the transport nodes are classified into four types (AG1,AG2, AC1, and AC2 * ), as described and illustrated in the * There is no AC2 in the T1 topology due to the rings. figure. On the one hand, this classification helps on havingflexibility in the choice of parameters (i.e., the values of thecharacteristics) for each topology and scenario. On the otherhand, the number of parameters is small (i.e., only four) incomparison to the number of transport nodes, e.g., 51 in thesmallest topology.
Aggregation Links Access LinksAG1 AC1 AC2AG2
Core Core
T1 T2
Fig. 5. Types of evaluated RAN topology.
Computing Resources.
We focused on the processingcapacity (i.e., CPU) because this metric has been the mostcommon bottleneck for the computing devices in the contextof network planning and vRAN optimization [25], [26]. CRvalues in Table 2 indicate the number of CPUs or a rangein the case of RC scenario. In this RC scenario, a specificvalue is random selected (inside the range) when gener-ating input data. The CRs capacity was designed takinginto consideration the CPU utilization profile of the RANsoftware from OpenAirInterface (OAI), as shown in Table3. The exact values may vary according to the adoptedsoftware components and computing device, but similarprofiles have been reported in different works.
Bandwidth.
The link capability considers the 5G-Crosshaul and PASSION projects and their strategies, i.e.,current and future networks, respectively. We define thebandwidth following the standards provided by the IEEEAlliance [29]. In this sense, we distinguish the interface ca-pacities between the two network topologies. 5G-Crosshauluses links capacity from 40 Gbps to 400 Gbps in the aggre-gation nodes (AG1 and AG2) and 10 Gbps to 40 Gbps in theaccess nodes (AC1 and AC2). PASSION Project (future net-work) operates from 100 Gbps to 1 Tbps in the aggregationnodes and from 40 Gbps to 100 Gbps in the access nodes.
Latency.
To reach the latency inputs, we considered fourcomponents. (i) Computing latency refers to the time con-sumed in the forwarding process. (ii) Fiber latency is relatedto propagation delay in optical fibers. (iii) The Optic is thedelay of the optical device without electronic processing. (iv)The regenerator transforms the optical signal into electricsignals. Based on these four components, we employed twostrategies, one for each topology. In the first strategy, T1uses only the Computing and Fiber components. However,all link distances in T1 are available, so the propagationdelay is directly computed by the distance versus delaypropagation. In the second strategy, T2 cannot directly
TABLE 2Scenarios employed in the evaluation.
Scenarios Low Capacity (LC) Random Capacity (RC) High Capacity (HC)Transport Nodes AG1 AG2 AC1 AC2 AG1 AG2 AC1 AC2 AG1 AG2 AC1 AC2Computing 1632 1632 816 816 32 32 16 16Resources †‡
16 16 8 8 1664 1664 832 832 64 64 32 32100 40 25 10 100400 40100 2540 1025 400 100 40 25Bandwidth (Gbps) †‡
800 100 50 40 1000800 400100 10050 5040 1000 400 100 50Computing † , ‡ † , ‡ ‡ ‡ † , ‡ - F1 and R1 - F1 and R1 - F1 and R1 TABLE 3RAN Protocol Stack CPU Utilization [26]–[28].
RAN Protocol CPU Utilization (cores)RRC 0.49PDCP 0.49High RLC 0.0245Low RLC 0.0245High MAC 0.343Low MAC 0.343High PHY 0.833Low PHY 2.352Total 4.9 compute the propagation delay, although all componentsare considered. In this case, the PASSION project performeda study that derived statistical information about potentialtopologies minimum, average, and maximum link distanceand the number of hops. Therefore, the link distance isdeveloped with minimum data for HC scenario, average,and maximum for LC scenario, and all of them for RCscenario [30].
RU Nodes.
To evaluate the scenarios (and topologies)with different demands, we considered two configurationsfor the number of RU nodes connected to the transportnodes: F1 (Fixed 1) – exactly one RU node is connectedto each transport node and R1 (Random 1) – zero or one(randomly chosen during input data generation) RU nodeis connected to each transport node. However, no RU nodeis connected to any AG1 in any topology scenario, as recom-mended by 5G-Crosshaul [31] and PASSION [32] projects.Table 4 summarizes the parameters of the functionalsplits employed in the evaluation. The priority of each DRC(used in the third stage of the PlaceRAN model) followsthe O-RAN Alliance specifications [23]. According to thesplit options, the maximum tolerated latency is defined foreach RAN sub-path: backhaul (Core-CU), midhaul (CU-DU), and fronthaul (DU-RU). Moreover, according to thesplit options, the minimum acceptable bandwidth is definedfor each RAN sub-path. However, we adopted different RFdevices for T1 and T2 topologies, which implied in differentbandwidths for each network, as shown in Table 4. In T1, weassumed the RF devices have the following characteristics:40 MHz bandwidth, 32 antenna ports, 8 MIMO layers, 216Physical Resource Blocks (PRBs), and 15 kHz subcarrierspacing per macro BS [33]. In T2, we assumed the RF devices have the following characteristics: 100 MHz bandwidth,32 antenna ports, 8 MIMO layers, 132 PRBs, and 60 kHzsubcarrier spacing per macro BS [33].We run all experiments in a Virtual Machine (VM) withUbuntu 18.04, 16 vCPUs, 1 TB RAM, and 40 GB of the virtualdisk. The VM is hosted in a server DELL PowerEdge M620with two Intel Xeon E5-2650 @ 2 GHz. We used Python2.7.17 and docplex 2.4.61 for implementing the PlaceRANmodel, and the solver used was IBM CPLEX 12.8.0. Thesource code and the input data used in the evaluation ispublicly available on Github * . We organize the evaluation of the PlaceRAN model intofour parts as described in the following. The first part of theevaluation examines the correlation between the minimiza-tion of CRs and the maximization of the aggregation level.The second part presents details about the DRC choicesmade by PlaceRAN. The third part investigates the aggre-gation process in more detail, mainly how computing andnetwork resources impact this process. The fourth part ofthe evaluation confirms the relevance of the three stages ofPlaceRAN. In this part, we also evaluate the impact of thenumber of paths while employing k-shortest paths. Finally,we summarize our main observations of these four parts ofthe evaluation in the end the section.
CRs and Aggregation level
According to Equation 1, the primary objective of PlaceRANis finding the best trade-off between the minimum amountof CRs ( Φ ) and the maximum amount of grouped RANVNFs, i.e., the aggregation level ( Φ ). In order to makethem comparable, we normalized the amount of CRs andthe aggregation level in a percentage scale. The percent-age of used CRs corresponds to the ratio of CRs runningany positive number of virtualized functions from the set F (cid:48) = { f , f , f , f , f , f , f } . In order to compute thepercentage of aggregation level, we assume that the highestachievable value of Φ is given by |F (cid:48) | × |B| , which repre-sents all virtualized functions running in a single CR.Fig. 6 shows the percentage of used CRs (X-axis) andthe percentage of aggregation level (Y-axis) of each solution * https://github.com/LABORA-INF-UFG/NG-RAN-model TABLE 4Parameters of the functional splits.
DRC Split Options Tolerated latency - one way (ms) 5G-Crosshaul bandwidth (Gbps) PASSION bandwidth (Gbps)N o Priority High Low Core-CU CU-DU DU-RU Core-CU CU-DU DU-RU Core-CU CU-DU DU-RU1 4 O1 O7 1.5 ∼
10 1.5 ∼
10 0.250 2.97 5.4 17.4 9.9 13.2 42.62 1 O2 O7 1.5 ∼
10 1.5 ∼
10 0.250 2.97 5.4 17.4 9.9 13.2 42.67 6 O1 O6 1.5 ∼
10 1.5 ∼
10 0.250 2.97 5.4 5.6 9.9 13.2 13.68 5 O2 O6 1.5 ∼
10 1.5 ∼
10 0.250 2.97 5.4 5.6 9.9 13.2 13.612 10 O1 - 1.5 ∼
10 1.5 ∼
10 - 2.97 5.4 - 9.9 13.2 -13 9 O2 - 1.5 ∼
10 1.5 ∼
10 - 2.97 5.4 - 9.9 13.2 -17 8 - O6 1.5 ∼
10 - 0.250 2.97 - 5.6 9.9 - 13.618 7 - O7 1.5 ∼
10 - 0.250 2.97 - 17.4 9.9 - 42.619 25 - - 1.5 ∼
10 - - 2.97 - - 9.9 - - obtained in three scenarios (LC, RC, and HC) and in twoconfigurations of RU nodes (F1 and R1). As expected, ascapacity increases there is a trend in increasing the aggre-gation level and decreasing the used CRs. However, theRAN topology has a significant impact on the solutions. T1topology, which represents the present RAN topology, tendsto limit the benefits of increasing the resources. For example,while comparing the low amount of resources (LC) with theintermediary amount of resources (RC), the configurationF1 exhibits improvement in aggregation level but not in thenumber of used CRs. Additionally, the hierarchical topologyadopted by T2 is robust to the configurations of RU nodes,i.e., to the demand, mainly as the amount of resourcesincreases. In general, the modern T2 topology also presentsbetter solutions in both aspects, percentage of aggregationlevel and percentage of used CRs. (a) T1 topology (b) T2 topology
Fig. 6. The relation between aggregation level and number of used CRs.
Fig. 6 represents the first stage of the PlaceRAN modeland the consequent totality of aggregation levels. For a moredetailed study, we show Table 5 considering the resultsof the Transport Nodes (AG1, AG2, and AC1) aggregationlevel and the CRs usage. However, we analyzed only the T1topology because this network presents non-homogeneousbehavior compared with the T2 topology. In this context,we did not consider the AC2 type of Transport Nodesbecause this does not found in T1 topology. Based on Table5, we can observe that AG1 concentrates aggregations of thevirtualized functions of 92.3% for the HC scenario with theF1 for RU Nodes and 79.6% for R1. Moreover, it aggregatedin the AG2 nodes only one CR using R1 and two CRswith F1, further than AC1 did not present any aggregationlevel. In this sense, it aligned with the model objectivesof concentrating VNFs near the core network. For the RCscenario, aggregations were less than 60% in AG1, andespecially for the Transport Nodes type F1 reached 30.8% in AC1. Finally, we observed a high level of aggregationin AC1, with 62.8% and 81.6% for the Transport Nodes R1and F1 in the LC scenario. For the LC and RC scenarios,we observe an increased use of CRs in AC1 nodes, whichresults in low aggregation of RAN’s VNFs.
TABLE 5Aggregation level of the T1 topology
LC RC HCR1 F1 R1 F1 R1 F1* 26.7 18.4 58.9 51 92.3 79.6AG1
Finding:
The PlaceRAN model achieves high aggrega-tion levels, e.g., above 80%, using only two CRs for AG1on the HC scenarios. However, we observed 0% of the ag-gregation level on specific Transport Nodes of AC1 type inthe same scenario. The LC and RC scenarios did not achievethe same aggregation level because of crosshaul restrictions(bandwidth and latency) and the ring format’s topology. Forthese scenarios, the PlaceRAN model reached the best ofnetwork achieving, e.g., the model obtained above 50% ofthe aggregation level on AG1 for the RC scenario.
DRCs options
This study analyzes the DRCs chosen by the PlaceRANmodel. Fig. 7 shows the amount and options of DRCs inthree scenarios (LC, RC, and HC) with the two RU Nodes(F1 and R1) for both investigated topologies. We join thenine DRCs from the five architectures shown in Fig. 1 infour groups for a precise analysis. In this context, we groupthe four DRCs (DRC1, DRC2, DRC7, and DRC8), namedNG-RAN (3), from 3 Independent Nodes architectures (O-RAN Low Split and SCF Low Split). We called NG-RAN(2) for DRC12 and DRC13 from the RU and DU integrationarchitecture. Moreover, we consider DRC17 and DRC18 inthe C-RAN architecture and DCR19 in the D-RAN archi-tecture. Furthermore, the shade of their respective color isaccording to the determined weights DRCs defined on thethird stage of the PlaceRAN model. In this case, the choiceof the best DCRs set is associated with the latency andbandwidth restrictions imposed by the network and the CRsavailable. (a) T1 topology (b) T2 topology
Fig. 7. DRCs chosen on the three scenarios
The first observation concerns the total number of DRCsassessed by each scenario, given the difference of the topolo-gies and strategies between RU Nodes R1 and F1. In thissense, Fig. 7(a) presents the T1 topology with 39 DRCs forR1 and 49 DRCs for F1. Moreover, Fig. 7(b) shows 101DRCs for R1 and 126 DRCs for F1 in the T2 topology.Observing the DRCs result’s behavior, Fig. 7 shows that thePlaceRAN model applied a distinct number of DRCs typesfor each scenario. For example, the model defined two typesof DRCs for the LC scenario with RU Nodes F1 for the T1topology and five DRCs types for the same scenario in theT2 topology. The DRC types increase is expected becauseof two main factors: (i) maximization of aggregation and(ii) crosshaul restrictions. For the first, the model’s priorityis to aggregate the maximum of VNFs following the DRCsweight of Table 2. Moreover, the model did not prioritizethe reduction of the number of DRCs types. For the secondfactor, the crosshaul restriction limited the range of thenumber of DRCs types chosen due to the crosshaul linkcapacity (bandwidth and latency). It is worth mentioningthat the CRs capacity applied not presented restriction forany scenarios.Another significant analysis is the choice of DRCs ac-cording to the architectures defined. For example, the Plac-eRAN model did not define D-RAN for thw HC scenario inboth topologies. However, D-RAN has high relevance in theT1 topology with the LC scenario, mainly in the LC scenariowith RU Node F1, reaching 40.8% (
Finding:
It is mandatory to deploy the D-RAN archi-tecture on the LC scenario in both topologies and on theRC scenario in only the T1 topology to reach the totalnumber of RU Nodes (R1 e F1). However, the HC scenarioachieved the combination of NG-RAN(3) and C-RAN above70% of the network’s deployment. Furthermore, the NG-RAN(3) architecture achieved 100% of the CU centralizationin the AG1 Transport Nodes for both topologies and the HCscenario. The exception was the T1 topology with the LCscenario, which did not support the NG-RAN(3) architec-ture requirements.
Network Resources Impact
This analysis investigates the correlation between the threestages of the PlaceRAN model and the network resourcesoccupation. Based on the two previous analyses, we observethat the crosshaul was the main restriction for the PlaceRANmodel. In this sense, we chose the T1 topology to high-light the constraints of this topology in the design of fifth-generation networks. Therefore, we selected ten characteris-tics for investigating this correlation for the three scenarios(LC, RC, and HC), with the two RU Nodes, as shown in Fig.8. We organize these characteristics into two types since theyare associated with the model and the network capacity. Thecharacteristics based on the PlaceRAN model are indicatedwith a red marker (aggregation level, DRCs usage, andDRCs weight). The network capacity is associated withTransport Nodes’ characteristics (AC1 link, AC1 CRs, AG2link, AG2 CRs, AG1 link, AG1 CRs, and latency), and wemark it in blue in Fig. 8. (a) RU Nodes R1
Latency (b) RU Nodes F1
Fig. 8. Model stages versus network resources evaluation.
Observing the characteristics of the PlaceRAN model,the RU Nodes R1 in Fig. 8(a) and F1 in Fig. 8(b) show similar results for the three scenarios. The aggregation levelpresented the same evaluating in the HC scenario, upperthan 60% and lower than 20% in the LC scenario. The DRCsusage showed the number of DRCs types in percent unit(nine DRCs equals 100%). The PlaceRAN model appliedalmost 60% of total DRCs types in the RC scenario andaround 20% in the LC scenario. In this case, we can see theappropriate aggregation level for the HC scenario and highDRCs usage for the RC scenario. Moreover, we can observethe DRCs weight, which is the sum result of DRCs chosenpriority detailed in Table 2. This characteristic shows whichsmaller outcome represents the best solution. Consequently,the HC scenario reached an effect lower of 20% with thebest solution. This behavior was the opposite for the LCscenario, with results around 60%. These results are con-nected directly with the network capacity, presented in thefollowing.The characteristics of Transport Nodes consumption,CRs, links, and latency are analyzing together becausethe behavior is complementary. These characteristics showthat the HC scenario had the highest occupation of RCsunder AG1 nodes, obtaining close to 70% for RU NodesR1 and 80% F1. Moreover, the latency average presenteda difference significant between the LC scenario and theHC scenario. For example, we observed a variation ofapproximately 45% for the RU Nodes R1 and 38% for F1.Considering absolute values, it is upper 5ms in the LCscenario, contributing directly to the low performance inthis scenario. For instance, the requirement of DU-RU was0.250ms, as presented in Table 2.The HC scenarios’ link capacity did not reach more than25% of occupation, resulting in a high aggregation of VNFsshowed by the aggregation level. The LC scenario presenteda high AC1 CRs and the link occupation, compared with theHC scenario because this topology has links that supportlow capacity. This behavior happens on the AG1 link occu-pation, for example, almost 60% of occupation, and the highdelay average. It is resulting in a low aggregation level andthe DRCs usage with high DRCs weight. However, the AG2link did not present a significant consumption variation forthe scenarios. This behavior occurs since the T1 topology hasa particular characteristic, where AG1 nodes concentrate all51 nodes and AG2 nodes only part of these nodes. Finding:
The crosshaul networks’ latency average mustbe compatible with the RAN demand (upper 5ms for the LCscenario) to achieve better performance of the T1 topology,i.e., based on a ring network. The capacities of the links andCRs are feasible for the optimization reduction. For exam-ple, the links and the CRs did not exceed 30% occupationfor both RU Nodes in the AG2 link. Otherwise, the AG1CRs occupation almost 80% for the HC scenario, showedthe best of the PlaceRAN model results.
Three stages of PlaceRAN model and k -paths In this last analysis, we investigate the PlaceRAN modelfrom two perspectives. First, we study the importance andbehavior of the three stages of the model. After, we observethe relation of K -paths with solutions found by the Plac-eRAN model. The second and third stages of the PlaceRANmodel deal with tie cases and define which solution is better.In this sense, Table 6 shows an example of a real case with the RC scenario for RU Nodes F1 in the T1 topology. Thesolutions found by the second and third stages presentssignificant improvements to the final solution. Therefore,Table 6 shows the nine DRCs configurations chosen withthe solutions found by three stages. The first stage definedthe maximum aggregation level and minimum number ofCRs operating in this scenario. The second stage found thesmallest number of DRCs that obtain the aggregation level,achieved in the first stage. It is fundamental to observe thatthe number of DRCs used in the second stage dropped fromeight to five. Finally, the third stage defined the picks ofDRCs according to the weights, which in this article, weprioritize the O-RAN solution. In this case, we can observethe number of DRC19 (D-RAN) decreases significantly, from . to . , given that D-RAN has low priority (weight)in O-RAN. TABLE 6Analyzing the three stages of the PlaceRAN model.
DRCs Stage 1 Stage 2 Stage 31 - NG-RAN(3) 2% 0% 0%2 - NG-RAN(3) 14.3% 22.4% 20.5%7 - NG-RAN(3) 0% 0% 0%8 - NG-RAN(3) 6.1% 10.2% 12.2%12 - NG-RAN(2) 2% 0% 0%13 - NG-RAN(2) 40.9% 42.9% 51%17 - C-RAN 8.2% 0% 0%18 - C-RAN 12.2% 12.2% 12.2%19 - D-RAN 14.3% 12.3% 4.1%Total 100% 100% 100%
Finally, we analyze the PlaceRAN model results for dif-ferent amounts of K -paths from the core to each RU Nodesspread across the topology. In some cases, PlaceRAN did notfind an optimal solution but a possible and viable solution,e.g., in some configurations with the T1 topology. Therefore,we calculated the objective function gap (%) between thispossible solution and the optimal one. For example, Fig. 9shows k -paths for the three scenarios with RU Nodes F1 inthe T1 topology. The x -axis represents the variation of k -paths between ( ≥ k ≤ ) since (six is gold Por que at´e 6?).Moreover, the y -axis depicts three views: (i) the objectivefunction value in the top, (ii) the amount of CRs used by themodel in the middle, and (iii) the aggregation level achievedby the solution in the bottom bars.We can observe that the PlaceRAN model reached apossible objective function near 60%, using 100% of CRs,and above 60% of aggregation level with k = 1 for the HCscenario. In the same analysis, the model found the objectivefunction near 40% for the RC scenario with K = 2 , alsousing 100% of CRs and 50% of the aggregation level. We cansee that the objective function, CRs, and aggregation levelstabilized for k ≥ . This stabilization means that for k > ,the model’s solutions did not present significant variationsfor the T1 topology because k > increases the problem’scomplexity by adding redundant paths in the topology. It isimportant to highlight that the same behavior occurs for theT2 topology. Finding:
The present evaluation proves that the threestages increase the model’s performance, for example, witha DRC2 growth (highest priority) by 6.2% and a decreasein DRC19 by 10.2% between stages 1 and 3 in T1 with F1 (a) T1 topology Fig. 9. K-shortest paths for the T1 topology. input. Another relevant analysis concerns the stabilizationof the model, which in T1 input F1, the paths K = 3 stabilizethe solution of the model, without the need to increase thenumber of paths, which leads to decreases in the problem’scomplexity in reason of redundant paths in the topology. Discussion
This section discusses the PlaceRAN model results based onthe fourth analysis with a global view aiming to correlatethe model’s characteristics. Achieving the model’s objectiveis necessary to overcome the crosshaul constraints for thebest placement of RAN’s VNFs to maximize the aggregationlevel, minimize the CRs, decide to the lower DRCs numbers,and prioritize the DRC choices based on strategies. In thissense, the solution takes the placement best decision focusedon overcoming the crosshaul limitations. However, the con-strain did not restrict the D-RAN architecture picked up insome scenarios, limiting the fifth-generation network’s de-ployment. The PlaceRAN model reached a high aggregationlevel for scenarios without crosshaul restriction resources,e.g., in the two AG1 CRs, centralizing the high RAN’s VNFsand leverage the fifth-generation networks based on O-RAN initiatives. Moreover, the PlaceRAN model found thebest optimize of k -paths possible considering the crosshaulcharacteristics.The behavior of the position of CU nodes and theiraggregation in the Transport Nodes (mainly in AG1 nodes)has significant importance for RAN development. For NG-RAN(3) architecture, the aggregation CU nodes VNFs areall in the AG1 nodes for both topologies, less for the LCscenario of the T1 topology, which does not have NG-RAN (3) DRCs types. This aggregation level presented asthe best option of deploy the vNG-RAN. However, the C-RAN architecture did not have the same performance once itachieves 100% of CU nodes VNFs in the HC scenario in theAG1 nodes and 100% of CU nodes VNFs in the LC scenarioin the AC1 nodes for the T1 topology. In this case, theaggregation in the AC1 nodes did not contribute to the CUcentralization. The NG-RAN(2) result presented a spreadaggregation of CU nodes in the different scenarios butreaches 0% of CU node position in the AC1 for any scenariosof the T2 topology and only 0% for the HC scenario in the T1 topology. Moreover, D-RAN did not have this behaviordue to the monolithic architecture.Finally, although the work does not explore the costsissue, the three scenarios chosen are directly oriented to in-vestment. For example, for the HC scenario with the best so-lution performance, there is a greater need for resources onthe network and consequently high investment. Moreover,the topologies analyzed show that the structure influencesthe results. For instance, the T1 topology is not fitted to thefifth generation networks’ evolution, mainly because of totallatency considering several computational processing of therouting devices in the AC1 layer. Furthermore, we analyzethe T1 topology with an intermediate RAN channel with40MHz, different from the 100MHz forecasted. Therefore,our analysis shows the need to redesign the network andthe investment for the T1 topology. In the same way, theT2 topology also needs investment, mainly because of theoptical devices for the crosshaul. Moreover, the projectanalyzed (5G-crosshaul and PASSION) show an imbalancebetween the occupation of resources. For example, in theHC scenario, AC1 nodes, and AC2 nodes presented a lowoccupation in links and CRs, which can be optimized, con-sequently reducing investment. ELATED W ORK
Recently, RAN has faced an intense process of softwarization and virtualization, which has driven into a fast evolu-tion. However, this scenario also created a misalignmentof several works related to the (virtual) network functionplacement due to the raise of multiple initiatives in a shortperiod of time. These initiatives are usually led by thestandard developer organizations, e.g., 3GPP, ITU-T, andETSI. Moreover, the O-RAN alliance drives some directions,such as vNG-RAN focusing on the disaggregation of radiofunctions for network efficiency and performance. In thiscontext, Table 7 shows several relevant works regardingOptimization Goals and Disaggregated RAN, consideringthe type of the nodes modeled and the number of DRCs.
Optimization Goals.
This characteristic is the most rele-vant when comparing the related work since investigationswith different optimization goals have distinct problem for-mulations, achieve different results, and, commonly, driveto diverse insights. In some works, the focus is to maximizethe number of VNFs running in a single CU. For example,for Garcia-Saavedra et al., the CU is co-located with the core,and D-RAN is not an option. However, for Fonseca, Correa,and Cardoso [15], D-RAN is taken into consideration, anddifferent positions are evaluated for CU. The maximizationof CU and BBU nodes are aimed at in several investigations[14], [15], [34], [35], [40]. In C-RAN architecture, the fron-thaul latency and the capacity of data rate links are widelyanalyzed [12], [16], [17], [19], [37], [40]. Based on RANcentralization, Song et al. [18] and Matoussi et al. [37], focuson the CRs efficiency. Furthermore, the RAN transformationis targeted into maximization of centralization, the CRsefficiency, and crosshaul link costs assessment by Arouk etal. [38] and Masoudi, Lisi, and Cavdar [39].Murti et al., in their initial work [12] and in its extendedversion [13], have the closest investigation to ours. Theseauthors are also approaching the problem of finding the TABLE 7Related work.
Optimization Disaggregated RANWorks Goals Nodes DRCs[14] Maximization of CU centralization and evaluation of Latency Edge CU-RU 3[15] Maximization of CU centralization and flexible CU positioning CU-RU 5[34] Maximization of BBU centralization and minimize the overall latency BBU-RRH 1[17] Dynamic CU and wireless fronthaul to minimize the energy CU-DU 4[19] Minimize the cost of VNF chain on substrate network BBU-RRH 1[35] Identify the ideal position of the centroid for the BBU BBU-RRH 1[36] Maximize DU distribution under non-dedicated optical network CU-DU 2[37] Minimize the CRs and fronthaul bandwidth BBU-RRH 3[18] Minimize the computational costs Monolithic 1[38] Minimize the CRs and overall routing costs CU-DU-RU 1[39] Minimize the total cost of ownership by DU pool DU-RU 4[16] Minimize the nodes and latency; maximize the data rate CU-DU-RU 2[40] Minimize interference and fronthaul links to optimize the network CU-DU 4[12], [13] Minimize the vRAN cost and overall routing, based on CU’s positioning CU-DU-(RU) 3PlaceRAN Minimize CRs and maximize vNG-RAN’s radio functions aggregation CU-DU-RU 9 best trade-off between the minimum number of CRs andmaximum aggregation level. However, they consider onlyvCUs, while DUs are fixed and close to (RU)s, i.e., providedas input. In our work, we consider not only vCUs, but alsovDUs, making our problem more general. We formulatedthe problem as a Nonlinear Programming model, with bi-nary variables, linear constraints, and a nonlinear objectivefunction (from the first stage). We were able to create anequivalent representation of the nonlinear objective functionusing Min and Max functions. Therefore, our model can besolved exactly by a conventional solver, e.g., the IBM CPLEX.
Disaggregated RAN.
The number of types of the RANnodes and the number of DRCs are related to the flexibilityand complexity of the problem under investigation. Thetypes of the RAN nodes available and the number of DRCsrepresented are related to the accuracy of describing thereal-world disaggregated RANs. In this context, our modelrepresents all types of the RAN nodes and all industryDRCs, i.e., our model describes precisely the present dis-aggregated RANs. Additionally, our model is ready forsupporting other DRCs by just changing the set D .During the classification of the related work in terms ofRAN nodes and number of DRCs, we adopted a conserva-tive approach. Therefore, most of the related work had itsbenefits amplified for the disaggregated RAN, e.g., Murtiet al. [12], [13] describe CU, DU, and RU in the systemmodel, but the paths between from RU are not taken intoaccount. This happens because each RU has only a singlelink connecting to a single DU. Therefore, our model is againmore general than the state-of-the-art.Given the complexity of the VNF placement problem,several works in the literature adopt an approximate andheuristic approach [17], [19], [34], [36], [37], [40]. As ex-pected, in general, the main advantage of this type ofapproach is the reduced computing cost and the maindrawback is the suboptimal solutions. As discussed in [13],suboptimal vRAN solutions have large cost impact in thelong-tem. Additionally, our strategy has advantages in com-parison with approximate and heuristic approaches. First,PlaceRAN is able to obtain the optimal solution in satisfac-tory time for several real-world networks, mainly for themost modern topologies. Second, when the computation isinterrupted before achieving the optimum, we know the gap of the suboptimal solution obtained by PlaceRAN. In this work, we address the optimization problem of vNG-RAN placement, which is still a paradigm for standard-ization and industry and relevant for deploying the fifth-generation RAN with the PlaceRAN solution. In this sense,we developed an optimal optimization solution to deliverthe best possibility of allocating RAN’s VNFs to reducethe use of computational resources and reach maximumaggregation and consequent centralization of RAN’s pro-tocols and units. In addition to overcoming the limitationsof crosshaul networks by choosing the best path. We de-veloped the concept of DRC disaggregation combinations,reduced the number of DRCs in the network, and insertedthe strategy concept for choosing DRCs. All this in scenariosand environments of real networks and with the possibilityof implementation in any vNG-RAN.For future work, we envision advances in the solutionconcerning the time factor. In the scope of choosing the bestpath for crosshaul networks, we will evolve the solutionto deal with traffic in the flow format, avoiding, evenmore, the waste of bandwidth. In computing resources, ourstrategy is to be even more aligned with O-RAN initiatives.Therefore, we will be introducing the type of processing tobe defined for each virtualized radio function, for example,general-purpose processor (GPP) or specific purpose pro-cessor (SPP). A CKNOWLEDGMENT
This work was conducted with partial financial supportfrom the National Council for Scientific and TechnologicalDevelopment (CNPq) under grant number 130555/2019-3and from the Coordination for the Improvement of HigherEducation Personnel (CAPES) - Finance Code 001, Brazil. R EFERENCES [1] 3GPP, “System Architecture for the 5G (Release 15),” TechnicalRecommendation 23.501, 2018.[2] G.-T. ITU-T, “Transport network support of IMT-2020/5G,” 2018. [3] T. ETSI, , 2019.[4] L. Gavrilovska, V. Rakovic, and D. Denkovski, “From Cloud RANto Open RAN,” Wireless Personal Communications , pp. 1–17, 2020.[5] P. Marsch et al. ,
5G system design: architectural and functional consid-erations and long term research . John Wiley & Sons, 2018.[6] R. Agrawal et al. , “Cloud RAN challenges and solutions,”
Annalsof Telecommunications , vol. 72, no. 7-8, pp. 387–400, 2017.[7] C. Badulescu and J. Triay, “ETSI NFV, the Pillar for Cloud ReadyICT Deployments,”
Journal of ICT Standardization , vol. 7, no. 2, pp.141–156, 2019.[8] C. J. Bernardos et al. , “Network virtualization research challenges,”IETF, Tech. Rep., 2019.[9] P. Sehier et al. , “Transport Evolution for the RAN of the Future,”
Journal of Optical Communications and Networking , vol. 11, no. 4, pp.B97–B108, 2019.[10] M. Masdari et al. , “An overview of virtual machine placementschemes in cloud computing,”
Journal of Network and ComputerApplications , vol. 66, pp. 106–127, 2016.[11] A. Laghrissi and T. Taleb, “A survey on the placement of virtualresources and virtual network functions,”
IEEE CommunicationsSurveys & Tutorials , vol. 21, no. 2, pp. 1409–1434, 2018.[12] F. W. Murti et al. , “On the Optimization of Multi-Cloud VirtualizedRadio Access Networks,” arXiv preprint arXiv:2002.10681 , 2020.[13] F. W. Murti et al. , “Optimal Deployment Framework for Multi-Cloud Virtualized Radio Access Networks,”
IEEE Transactions onWireless Communications , pp. 1–1, 2020.[14] A. Garcia-Saavedra et al. , “Fluidran: Optimized vRAN/MEC or-chestration,” in
IEEE INFOCOM Conference on Computer Communi-cations , 2018, pp. 2366–2374.[15] F. Fonseca, S. Correa, and K. Cardoso, “Optimizing allocationand positioning in a disaggregated radio access network awareof paths through the core infrastructure,” in
Anais do XXXVIISimp´osio Brasileiro de Redes de Computadores e Sistemas Distribu´ıdos ,2019, pp. 791–804.[16] J. Yusupov et al. , “Multi-objective function splitting and placementof network slices in 5G mobile networks,” in
IEEE Conference onStandards for Communications and Networking (CSCN) , 2018, pp. 1–6.[17] D. Harutyunyan et al. , “CU placement over a reconfigurable wire-less fronthaul in 5G networks with functional splits,”
InternationalJournal of Network Management , vol. 30, no. 1, p. e2086, 2020.[18] S. Song et al. , “Clustered virtualized network functions resourceallocation based on context-aware grouping in 5G edge networks,”
IEEE Transactions on Mobile Computing , vol. 19, no. 5, pp. 1072–1083, 2019.[19] O. Arouk et al. , “Multi-objective placement of virtual networkfunction chains in 5G,” in
IEEE 6th International Conference on CloudNetworking (CloudNet) , 2017, pp. 1–6.[20] L. M. P. Larsen, A. Checko, and H. L. Christiansen, “A Surveyof the Functional Splits Proposed for 5G Mobile Crosshaul Net-works,”
IEEE Communications Surveys & Tutorials , vol. 21, no. 1,pp. 146–172, 2019.[21] 3GPP, “Study on New Radio Access Technology; Radio Access Ar-chitecture and Interfaces (Release 14),” 3 r d Generation PartnershipProject (3GPP), Technical Recommendation (TR) 38.801, 2017.[22] Mavenir, “Transforming the Radio Access Network - Mavenir’sVirtualized RAN,” Mavenir, Tech. Rep., 2019.[23] O. R. Alliance, “Cloud Architecture and Deployment Scenarios forO-RAN Virtualized RAN,”
White Paper , 2020.[24] 3GPP-TR21.916, “System Architecture for the 5G (Release 16),”Technical Recommendation 21.916, 2020.[25] P.-C. Lin and S.-L. Huang, “Performance profiling of cloud radioaccess networks using openairinterface,” in
Asia-Pacific Signal andInformation Processing Association Annual Summit and Conference(APSIPA ASC) , 2018, pp. 454–458.[26] Y. Hu and J. Wang, “Architectural and Cost Implications of the5G Edge NFV Systems,” in
IEEE 37th International Conference onComputer Design (ICCD) , 2019, pp. 594–603.[27] C. Y. Yeoh et al. , “Performance study of LTE experimental testbedusing OpenAirInterface,” in et al. , “Efficient virtual network function placementstrategies for cloud radio access networks,”
Computer Communica-tions , vol. 127, pp. 50–60, 2018.[35] B. Mahapatra et al. , “Optimal Placement of Centralized BBU (C-BBU) for Fronthaul and Backhaul Optimization in Cloud-RANNetwork,” in
International Conference on Information Technology(ICIT) , 2017, pp. 107–112.[36] N. Molner et al. , “Optimization of an integrated fron-thaul/backhaul network under path and delay constraints,”
AdHoc Networks , vol. 83, pp. 41–54, 2019.[37] S. Matoussi et al. , “5G RAN: Functional Split Orchestration Opti-mization,”
IEEE Journal on Selected Areas in Communications , vol. 38,no. 7, pp. 1448–1463, 2020.[38] O. Arouk et al. , “Cost optimization of cloud-RAN planning andprovisioning for 5G networks,” in
IEEE International Conference onCommunications (ICC) , 2018, pp. 1–6.[39] M. Masoudi, S. S. Lisi, and C. Cavdar, “Cost-Effective MigrationToward Virtualized C-RAN With Scalable Fronthaul Design,”
IEEE Systems Journal , 2020.[40] D. Harutyunyan and R. Riggio, “Flex5G: Flexible functional splitin 5G networks,”