Latency Bounds of Packet-Based Fronthaul for Cloud-RAN with Functionality Split
LLatency Bounds of Packet-Based Fronthaul forCloud-RAN with Functionality Split
Ghizlane Mountaser, Maliheh Mahlouji, Toktam Mahmoodi
Centre for Telecommunications ResearchDepartment of Informatics, King’s College LondonLondon WC2B 4BG, UK { ghizlane.mountaser, maliheh.mahlouji, toktam.mahmoodi } @kcl.ac.uk Abstract —The emerging Cloud-RAN architecture within thefifth generation (5G) of wireless networks plays a vital role inenabling higher flexibility and granularity. On the other hand,Cloud-RAN architecture introduces an additional link betweenthe central, cloudified unit and the distributed radio unit, namelyfronthaul (FH). Therefore, the foreseen reliability and latency for5G services should also be provisioned over the FH link. In thispaper, focusing on Ethernet as FH, we present a reliable packet-based FH communication and demonstrate the upper and lowerbounds of latency that can be offered. These bounds yield insightsinto the trade-off between reliability and latency, and enable thearchitecture design through choice of splitting point, focusingon high layer split between PDCP and RLC and low layersplit between MAC and PHY, under different FH bandwidthand traffic properties. Presented model is then analyzed bothnumerically and through simulation, with two classes of 5Gservices that are ultra reliable low latency (URLL) and enhancedmobile broadband (eMBB).
Index Terms —Cloud-RAN; Fronthaul; Ethernet; Latency; Re-liability; Upper bound; Lower bound.
I. I
NTRODUCTION
One of the architecture enablers of fifth generation (5G) isCloud-RAN that is supported through multiple technologicaladvances in the network including softwarization, virtual-ization and cloudification [1]. Cloud-RAN despite bringinghigher flexibility and granularity to the network architecture,introduces an additional communication link, i.e. fronthaul(FH). Hence, the ability for the FH to flexibly scale upwith data rate has become critical to the success of Cloud-RAN. The need for flexibility in the FH has opened up thepossibility of flexibly splitting Radio Access Network (RAN)functionalities between central unit (CU) and distributed unit(DU). The advantage to such an architectural approach is theuse of different transport such as packet-based FH. Adoptingpacket-based FH in the Cloud-RAN architecture allows the useof widely deployed Ethernet-based network. At the same time,packet-based networks impose challenges in ensuring the highreliability and low-latency over the FH communication, whichare the key performance indicators expected of 5G.Reliability can be typically achieved by retransmission orredundancy. However, reliability is usually increased at thecost of latency which poses a major challenges when latencyrequirement is very stringent. For this reason, the design ofCloud-RAN solution involves one key design question, that is, which functional splits may be suitable from a reliability-latency point of view under the constraint required by 5Gscenarios.Given that each split comes with its own delay requirements,having the knowledge of latency bounds of FH will allow us todecide which split is the most appropriate in the Cloud-RANarchitecture. On the other hand, 5G traffic classes eMBB (en-hanced mobile broadband), URLLC (ultra reliable low latencycommunications) and Massive machine type communications,each come with varied requirements on reliability that shouldalso be maintained on the FH link; improving reliability oftenresults in increasing latency. To this end, focusing on reliablepacket-based FH [2], the aim of this paper is to compute thelower bound and upper bound of latency analytically, usingstochastic network calculus [3], [4], [5]. We simulate thereliable packet-based FH, demonstrating where and how delaybounds are achieved for two classes of traffic, which are eMBBand URLLC. Having these bounds, we further analyse whereeach functionality split can be the best architectural choice.This paper is an expended work to our previous work [2] wherewe investigated how to improve reliability and latency ofpacket based fronthauling by means of multi-path diversity anderasure coding; and [6], [7] which examined lower layer andhigher layer splits through an experimental testbed consideringan Ethernet-based FH.The remainder of this paper is organized as follows. SectionII gives a brief overview on Cloud-RAN and its differenttransport technologies and explores reliability on the FH. Insection III, we elaborate system model of Cloud-RAN withmulti-path FH using coding to analyse reliability-latency andwe shed light on functional split requirements in term oflatency. In section IV, we compute stochastic delay lowerand upper bounds of the system model. The analytic andsimulation results are studied in section V. Finally, conclusionand future research are presented in section VI.II. B
ACKGROUND
Cloud-RAN is considered one of the key enablers of 5Garchitecture, given its desired properties [8]. Despite theattractive advantages of the conventional Cloud-RAN, thearchitecture whereby a standard common public radio interface(CPRI) is used to transport base band radio samples between a r X i v : . [ c s . N I] A p r U and DU faces several challenges. The first challenge isthe user data is transmitted in the form of an IQ-data blockwhich requires large bandwidth of . Gbps, considering a100 MHz transmission bandwidth and 32 antenna ports [9].Thereby, the high-throughput requirement poses challengesfor the FH interface. The second challenge is that CPRIrequires stringent latency and jitter requirements. It requiresan end-to-end latency of around µ s [10]. Such criticalrequirements are making CPRI very challenging. To relax theexcessive bandwidth and latency requirements, as well as toenhance the flexibility of the FH, functional split is introducedwhereby a more flexible placement of baseband functionalitybetween the DU and the CU is considered. In fact severaloptions of functional splits have been standardized in 3GPP[9]. Moreover, fundamental simulations and experimentationsare carried out by various academic studies [11], [7], [12].Nonetheless, each split point has different requirements suchas latency and bandwidth. These requirements have to beconsidered in order to select the appropriate functional split.In general, the lower the split point, the greater the level ofcentralization, the higher is the required interface data rate andthe more stringent is the latency requirement.In traditional Cloud-RAN architecture, fibre has been de-fined as an ideal attractive solution to meet the strict require-ments of high bandwidth and low latency of CPRI. However,there are situations where deployment of fibre is difficult ornot a good choice due to cost. In this end, packet-based FHcan be considered as a promising alternative transport. Thishighly cost effective solution allows sharing and convergencewith Ethernet-based fixed networks and offers great flexibility.However, packet-based FH imposes many challenges such ashigh latency and high jitter. Nevertheless, it can be used infunctional split where the latency and jitter requirements arerelaxed. For example, some of the authors of this work hasdemonstrated the feasibility of splitting between MAC andPHY in their past work [6]. Further studies have shown thatthe requirements of different 5G service classes, including theURLLC service can be accommodated using packet-based FH[7]. In [13] authors analysed impact of packetization on theCloud-RAN and they analyzed different packet scheduling toincrease the multiplexing gain.In addition to low latency and excessive data rate, reliabilityis also an important metric in 5G [14], [15]. Under theCloud-RAN architecture, FH needs to provide comparativereliability to enable adoption of Cloud-RAN. The most wellused methods to improve reliability are retransmission, multi-path with packet duplication and multi-path with coding.Retransmission is a straight forward way to achieve relia-bility. However, retransmission can have significant impacton increasing the latency making it non-viable solution onthe FH where delays cannot be afforded. By contrast, pathdiversity with duplication offers better latency in the expenseof significant transmission overhead by duplicating packetsover multiple interfaces increasing FH network congestion.Two important considerations in such approaches are latencyand FH overhead. An alternative solution that provides trade- off between latency and FH overhead is channel coding whichcan add controlled redundancy to achieve desired reliabilityand splits the total amount of information to transmit acrossdifferent paths. Additional reliability, using any technique,sacrifies latency and hence looking at the boundaries of latencythat can be offered under certain reliability is of interest to theapplication in-need of both [14].III. S YSTEM M ODEL
In this section, we first introduce the system model for theCloud-RAN system with multi-path FH and then shed a lighton functional split requirements in term of latency.
A. System Model
Our system model consists of Cloud-RAN with a single CUand a single DU connected with multiple FH paths ( n differentpaths), where each path i has a capacity ψ i . Packets of size B bits are arrived to the system with exponential inter-arrivalperiods with average /λ seconds (s). We assume that the FHlinks are identical. Each link is modelled as a single queue.We suppose that the service time of each queue follows anexponential distribution. The mean service time to transmit apacket of size B bits from CU to DU is /µ = B/ψ s. Thepackets within each queue is served in a first in first out mannerand the buffer length is assumed to be infinite. The focus of ourmodel is on downlink (DL) direction. However, all argumentsare valid in the reverse direction of communications.We analyze the performance of the system by consideringcoexistence of both eMBB and URLLC traffics over orthogo-nal and non-orthogonal sharing of FH resources as describedin [2] using multi-path FH with coding (MPC).In this solution (Fig. 1), packets arrive to the CU withexponential inter-arrival periods with average /λ s. Eachpacket goes through the four steps below, • Fragmentation block fragments the arrival packet into k equal blocks, each with size B/k . • Encoder block encodes the blocks into n encoded blockswith size B/k . Each block is then forked into n paths andserviced in parallel. The service time of each path followsan exponential distribution with service rate µ MPC = kψB . • At the receiver, the original packet can be retrieved ifany k out of n are received successfully. Thereby, once k blocks are received, they are passed into decoder to startdecoding them without waiting for the remaining ones. • The k decoded blocks are passed to concatenation blockto be merged into one packet.Latency in this solution is determined from the time packetis transmitted over the FH until k blocks are successfullyreceived. B. Functional Split and Latency Requirement
Different functional split points have different latency andbandwidth requirements on the FH [9]. These requirementsshould be considered to support 5G scenarios since eachscenario requires different end-to-end requirements in term oflatency and reliability as shown in Table I [16]. Hence, a split ath 1path n μ λ Fragmentation
Encoder k Blocks n BlocksB/K CU B Decoder k BlocksB/K B Concatenation DU μ MPCMPC
B/K .. ..
FH Network ...
URLLC UEeMBB UE
Fig. 1. Multi-path FH with erasure coding (MPC) for downlink communication.TABLE IR
EQUIREMENTS FOR
5G S
CENARIOS
Scenario End-to-end latency Reliability Payload sizeTactile interaction . ms . SmallElectricity distribution (high voltage) ms . SmallElectricity distribution (medium voltage) ms . Small to bigDiscrete automation ms . Small to bigIntelligent transport systems ms 99.9999% Small to bigTABLE IIB ANDWIDTH AND LATENCY REQUIREMENTS FOR DIFFERENT SPLITPOINTS
Split Point One-way Latency DL Bandwidth UL BandwidthPDCP-RLC . − ms Mbps
MbpsMAC-PHY µ s Mbps
Mbps point that can sustain scenario requirements can be consideredappropriate.Among all available splits, we will focus our attention onPDCP-RLC and MAC-PHY splits (option 2 and option 6respectively according to [9]). The expected latency require-ments as estimated by 3GPP [9] for each split is listed in TableII. • PDCP-RLC split: For this split, RRC and PDCP arecentralized whereas RLC, MAC, PHY and RF are dis-tributed. From a latency point of view, PDCP-RLC splithas a relaxed latency requirement on the FH. It tolerateshigh latency as PDCP doesn’t require a strict lower layersynchronization. The maximum tolerable one way latencyshould be in maximum ms. • MAC-PHY split: For this option, the split is betweenMAC and PHY wherein only PHY and RF are in DU. Thesplit offers a high level of centralization and pooling gaincompared to PDCP-RLC split. In this split, the HARQprocess and other timing critical functions are located inCU which results in tighter latency constraints on the FH.This split can support µ s latency in maximum.IV. S TOCHASTIC D ELAY B OUNDS FOR ( N , K ) FORK - JOIN S YSTEM
The fronthaul delay for the MPC method described insection III can be computed by analysing ( n, k ) fork-join system. Although ( n, n ) fork-join system, also known asbasic fork-join system, has been thoroughly studied, thereare many open problems in analysing its generalization, i.e. ( n, k ) fork-join system. Mean value analysis for ( n, k ) fork-join system has been done in [4] and [5]. Authors in [3]used stochastic network calculus to define a stochastic upperbound for distribution of delay in ( n, k ) fork-join system ina general case. Nevertheless, there is no reasonable way touse that formula without knowing the joint distribution ofparallel queues (in case of dependent queues). In this paper,we compute an upper bound and lower bound for ( n, k ) fork-join system delay using the concept of independency and fulldependency between parallel links.It is worth mentioning that compared to the work presentedin [5], we make an additional assumption of non-purgingscenario i.e. after k out of n blocks exit the queuing system, theother n − k remaining blocks are not removed from the queuesand they will continue being processed. This assumption ismore realistic in this context given dispatched packets cannot be removed from the links and switches. As it has beendiscussed in [5], split-merge system (which is a variation of ( n, k ) fork-join system that blocks processing of the nextpackets until k out of n blocks of the current packet finishbeing processed) provides an upper bound for delay in purgingscenario, however, it is not an upper bound in non-purgingscenario.In this section, our objective is to compute stochastic delaybounds of ( n, k ) fork-join system. As detailed in section III,in the MPC method, CU encodes the packet into n equallength blocks and sends those blocks into n parallel links.Hence assumption of independence between n links is notvalid. Moreover, without knowing the dependency betweenthe links, e.g. their joint distribution, computation of delaybounds are not tractable. Therefore, in this paper we calculatestochastic lower bound and upper bound for delay distributionnder certain assumptions. A. Stochastic Lower Bound for Delay
In a homogeneous ( n, k ) fork-join system, the smallest de-lay stochastically occurs when all links are independent, sincein that case if some links are highly congested, other linksmight be less congested with larger probability. Therefore,here we will find the stochastic distribution of delay in thecase that all links are independent. Authors in [3] computeddelay bound for independent links, using stochastic networkcalculus for a general case. However, we will derive this boundfor the case in which each parallel link is an M/M/ queueby applying classical queuing theory.For an M/M/ queue with the iid Poisson arrival process,with mean λ and iid exponentially distributed service timeswith mean /µ , we have P { d > τ } = e − ( µ − λ ) τ =: p , (1)where d is the block delay in an M/M/ queue whichincludes waiting time of the block in the queue plus its ownservice time. Let us assume ( n, k ) fork-join system, whichconsists of n parallel homogeneous M/M/ queues. In thissystem, delay of a packet is defined as the time betweenexecution of the n encoded blocks into the n parallel links untilthe first k out of n blocks have been processed in the queues.Note that Equation (1) can be viewed as a Bernoulli processin which, the number of successes in n independent trialshas Binomial distribution. Therefore, ( n, k ) fork-join delay,denoted by D , would be, P { D > τ } = k − (cid:88) j =0 (cid:18) nj (cid:19) (1 − p ) j p n − j . (2)which is the probability that more than n − k links havegreater delay than τ . B. Stochastic Upper Bound for Delay
Similarly in a homogeneous ( n, k ) fork-join system, thelargest delay stochastically occurs when all links are fullydependant, such that all queues have the same length; inthis case congestion happens at the same time in all links.Therefore, we use the “equal queue length” assumption tocompute the worst case of dependency, instead of looking forjoint distribution of the queues, and find a stochastic upperbound for ( n, k ) fork-join system delay.For an M/M/ queue, the queue length, denoted by L q ,would be equal to l with the following probability, P { L q = l } = (1 − ρ ) ρ l (3) ρ := λµ . (4)Also, delay profile for an M/M/ queue with length l isas follows, P { d > τ } = l (cid:88) m =0 ( µτ ) m m ! e − µτ := p . (5)Similar to the previous computation, to find the delay distri-bution of ( n, k ) fork-join system consisting of n homogeneous M/M/ queues, we should compute the probability that morethan n − k queues have the delay greater than τ . Therefore,in the case of dependant parallel queues, the ( n, k ) fork-joinsystem delay will be as follows, P { D > τ } = k − (cid:88) j =0 (cid:18) nj (cid:19) × P { d < τ, ..., d j < τ, d j +1 > τ, ..., d n > τ } , (6)where d i , i = 1 , ..., n denotes the block delay in the i th queue (i.e. i th link). In this analysis, we assume all parallellinks have the same queue length equal to L q . We furtherassume that solely this property, i.e. equal queue length,defines the dependency between links, while the queues areassumed to be independent. Hence, joint probability in Eq.(6) can be computed using Bayes’ law, as follows, P { D > τ } = k − (cid:88) j =0 (cid:18) nj (cid:19) × ∞ (cid:88) l =0 P { d < τ, ..., d n > τ | L q = l } P { L q = l } = k − (cid:88) j =0 (cid:18) nj (cid:19) × ∞ (cid:88) l =0 P { d < τ | L q = l } ...P { d n > τ | L q = l } P { L q = l } = k − (cid:88) j =0 (cid:18) nj (cid:19) ∞ (cid:88) l =0 (1 − p ) j p n − j (1 − ρ ) ρ l . (7)V. S IMULATION R ESULTS
In this section, we develop a simulation model in MATLABto validate our analysis in the presence of coexisting eMBBand URLLC services. Characterization of the two services areshown in Table III. We assume there are n = 10 independentFH paths, where each path i has a capacity of Mbps, i.e. ψ i = 100 Mbps, ∀ i .We initially plot the non-orthogonal sharing of FH resourceswith orthogonal FH transmission schemes that can allocatea different amount of resources to URLLC. In these firstplots, the aim is to determine the allocations that improvethe probability of error for a given latency for the URLLCservices in orthogonal as compared to non-orthogonal FHshared resources.In Fig. 2, we plot the error probability for URLLC usingorthogonal bandwidth allocation on the FH with differentURLLC bandwidth fractions; in each case bw u fraction of theavailable path bandwidth, ψ i , is allocated to URLLC. The plotshows that the choice bw u ≥ / ψ i can reduce the latencyas compared to shared FH transport. ABLE IIIS
YSTEM PARAMETERS FOR THE
5G S
ERVICES
Type of traffic eMBB URLLCPacket Size (Bytes) 1500 500 λ (packet/ms) 4 8Fig. 2. Achievable probability of error Vs. latency under orthogonal FHbandwidth with different bandwidth fractions for URLLC. Fig. 3 shows the error probability for URLLC using or-thogonal path allocation on the FH with different number ofpaths allocated to URLLC, i.e. using n u of the available path, n = 10 . The plot shows the latency to achieve the errorprobability of better than − can be reduced as comparedto shared FH transport by choosing n u ≥ . For examplethe error probability of − obtained using orthogonal pathsis improved by approximately as compared to thatobtained by non-orthogonal sharing of FH resources.Focusing on orthogonal bandwidth allocation on the FHwith bw u = 1 / ψ i , Fig. 4 shows that simulation resultsare bounded by the lower and upper bounds computed fromEquations (2) and (7). For URLLC (Fig. 4(a)), to achieve areliability of . the latency ranges from . ms to . ms. As for eMBB (Fig. 4(b)), the latency range is widervarying from . ms to . ms. From Fig. 5 we can observethe performance of both URLLC and eMBB are enhanced ascompared to orthogonal bandwidth allocation (Fig. 4).Focusing on Fig. 5, we use the lower and upper boundsobtained in this figure to choose appropriate functional splitthat offers the required reliability for a given scenario. Forexample using MAC-PHY split with the URLLC traffic (Fig.5(a)), the upper bound can provide a reliability of . at latency of . ms. Therefore, this setup can be used forlow latency applications which requires a reliability as high as . . Considering the requirements listed in Table I, thissetup can, for example, be used for all scenarios. As for eMBB(Fig. 5(b)), MPC with MAC-PHY split can offer a reliabilityless than . which is not suitable for any scenario listedin Table I whereby the reliability requirements are of at least . . In such a case MPC with PDCP-RLC split is the only Fig. 3. Achievable probability of error Vs. latency under orthogonal FH pathwith different number of paths for URLLC. choice available since the upper bound can provide a reliabilityof . at latency of . ms.To summarise, MAC-PHY split is the most appropriate splitfor scenarios using URLLC traffic considering the systemmodel and traffic patterns in Table III, since it meets theirlatency and reliability requirements. Whereas, PDCP-RLCsplit is more suitable for scenarios using eMBB traffic.VI. C ONCLUSION AND F UTURE R ESEARCH
In this paper, we presented a Cloud-RAN model basedon multi-path FH with coding solution for enhancing thereliability of the FH. The paper aims at providing an upperand a lower bounds of reliability-latency function on the FHunder orthogonal FH allocation.We first derived lower and upper bounds analytically. Thenwe simulated the Cloud-RAN model to demonstrate the ef-fectiveness of the analytic by showing the simulation resultsare bounded by lower and upper bounds. Finally, based onthis result, we discussed the recommendations for split pointfocussing on MAC-PHY and PDCP-RLC splits for differentscenarios to meet their latency and reliability requirements.Future work will be focusing on analyzing multi-path FHwith multi-hops. A
CKNOWLEDGEMENT
This work has been supported by The Engineering and PhysicalSciences Research Council (EPSRC) industrial Cooperative Awardsin Science & Technology (iCASE) award and by the British Telecom(BT). Additional support is received from EU H2020 5GCAR. R EFERENCES[1] M. Condoluci and T. Mahmoodi, “Softwarization and virtualization in 5gmobile networks: Benefits, trends and challenges,”
Computer Networks ,vol. 146, pp. 65 – 84, 2018.[2] G. Mountaser, T. Mahmoodi, and O. Simeone, “Reliable and Low-Latency Fronthaul for Tactile Internet Applications,”
IEEE JSAC ,vol. 36, pp. 2455–2463, Nov 2018.[3] M. Fidler and Y. Jiang, “Non-asymptotic delay bounds for (k, l) fork-join systems and multi-stage fork-join networks,” in
IEEE INFOCOM2016 - The 35th Annual IEEE International Conference on ComputerCommunications , pp. 1–9, Apr 2016.
Latency (ms) -6 -5 -4 -3 -2 -1 P r obab ili t y o f E rr o r MPC, k=2 (Simulation)Lower Bound (Analysis)Upper Bound (Analysis)
PDCP-RLCMAC-PHY (a) URLLC.
Latency (ms) -6 -5 -4 -3 -2 -1 P r obab ili t y o f E rr o r MPC, k=2 (Simulation)Lower Bound (Analysis)Upper Bound (Analysis)
PDCP-RLCMAC-PHY (b) eMBB.Fig. 4. Achievable probability of error Vs. latency at the upper and lower latency bounds, under bandwidth orthogonal FH with URLLC bandwidth fraction / ( bw e = 1 / ψ i and bw u = 1 / ψ i ). The dashed line represents the maximum latency supported by functional split
Latency (ms) -6 -5 -4 -3 -2 -1 P r obab ili t y o f E rr o r MPC, k=2 (Simulation)Lower Bound (Analysis)Upper Bound (Analysis)
MAC-PHY PDCP-RLC (a) URLLC.
Latency (ms) -6 -5 -4 -3 -2 -1 P r obab ili t y o f E rr o r MPC, k=2 (Simulation)Lower Bound (Analysis)Upper Bound (Analysis)
PDCP-RLCMAC-PHY (b) eMBB.Fig. 5. Achievable probability of error Vs. latency at the upper and lower latency bounds, under path orthogonal FH with URLLC path fraction / ( n e = 5 and n u = 5 ). The dashed line represents the maximum latency supported by functional split [4] H. Wang, J. Li, Z. Shen, and Y. Zhou, “Approximations and Bounds for(n, k) Fork-Join Queues: A Linear Transformation Approach,” in , vol. 00, pp. 422–431, May 2018.[5] G. Joshi, Y. Liu, and E. Soljanin, “Coding for fast content download,”in , pp. 326–333, Oct 2012.[6] G. Mountaser, M. L. Rosas, T. Mahmoodi, and M. Dohler, “On theFeasibility of MAC and PHY Split in Cloud RAN,” in
IEEE WCNC ,March 2017.[7] G. Mountaser, M. Condoluci, T. Mahmoodi, M. Dohler, and I. Mings,“Cloud-RAN in Support of URLLC,” in , Dec 2017.[8] China Mobile, “C-RAN: the Road Towards Green RAN,”
White Paper ,vol. 2, 2011.[9] 3GPP, “Study on New Radio Access Technology: Radio Access Archi-tecture and Interface (Release 14),” Tech. Rep. 38.801, 2017.[10] NGMN, “Further Studies on Critical Cloud RAN Technologies.” WhitePaper, March 2015.[11] A. S. Thyagaturu, Z. Alharbi, and M. Reisslein, “R-fft: Function split at ifft/fft in unified lte cran and cable access network,”
IEEE Transactionson Broadcasting , vol. PP, no. 99, pp. 1–18, 2018.[12] J. Baranda, J. Mangues-Bafalluy, I. Pascual, J. Nunez-Martinez, J. L. D.l. Cruz, R. Casellas, R. Vilalta, J. X. Salvat, and C. Turyagyenda, “Or-chestration of End-to-End Network Services in the 5G-Crosshaul Multi-Domain Multi-Technology Transport Network,”
IEEE CommunicationsMagazine , vol. 56, pp. 184–191, July 2018.[13] C. Chang, N. Nikaein, and T. Spyropoulos, “Impact of Packetizationand Scheduling on C-RAN Fronthaul Performance,” in , pp. 1–7, Dec 2016.[14] J. Sachs, G. Wikstrom, T. Dudda, R. Baldemair, and K. Kittichokechai,“5G Radio Network Design for Ultra-Reliable Low-Latency Communi-cation,”
IEEE Network , vol. 32, pp. 24–31, March 2018.[15] N. I. Sulieman, E. Balevi, K. Davaslioglu, and R. D. Gitlin, “Diversityand network coded 5G fronthaul wireless networks for ultra reliable andlow latency communications,” in