JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 38, NO.

6, MARCH 15, 2020 1125

Isolation-Aware 5G RAN Slice Mapping Over WDM

Metro-Aggregation Networks
Hao Yu , Student Member, IEEE, Francesco Musumeci , Member, IEEE, Jiawei Zhang ,
Massimo Tornatore , Senior Member, IEEE, and Yuefeng Ji , Senior Member, IEEE

Abstract—To accommodate the heterogeneous requirements of I. INTRODUCTION

emerging 5G services, 5G Radio Access Network (RAN) will sup-
LARGE number of new mobile services is emerging in
port the rising paradigm of network slicing, where a single physical
RAN infrastructure is divided into multiple logical networks char-
acterized by diverse requirements in terms of performance, cost,
A 5G mobile networks, such as Augmented/Virtual Reality
(AR/VR), automonous driving, industrial automation, etc. These
latency, and security, etc. The optimality of RAN slicing depends new services are characterized by heterogeneous requirements
on various factors, such as slice requirements, resource availabil-
ity and adopted network architecture. In 4G, RAN architecture
in terms of data rate, end-to-end latency and reliability [1]
evolved from a distributed architecture towards a 2-layer archi- and can be broadly classified into three categories: enhanced
tecture that separates remote radio heads (RRHs) from baseband Mobile Broadband (eMBB), ultra-Reliable Low-Latency Com-
processing units (BBUs). However, this architecture has soon shown munication (uRLLC), massive Machine Type Communication
scalability limitations. To address this problem, the concept of flexi- (mMTC). eMBB services represent the high-bandwidth de-
ble 5G functional split has been proposed, where RAN functions are
distributed on a 3-layer architecture featuring a radio unit (RU),
manding applications, e.g., AR/VR and live video. uRLLC
a distributed unit (DU), and a central unit (CU). Besides RAN services are characterized by a strict requirement on real-time
architecture, isolation among different slices, which is necessary interaction, e.g., as in autonomous driving. mMTC services in-
for slicing security, has a significant influence on the RAN slicing. clude both long-range MTC and broadband MTC with low-cost
In this work, we consider the isolation-aware RAN slice mapping and low-power, e.g., smart services for metering, environment
problem considering advanced functional splits and a 3-layer RAN
architecture. We propose a dual-objective heuristic algorithm for
monitoring and traffic control in the urban and rural areas,
RAN slice mapping over a physical infrastructure represented which result in very high device density. Satisfying these diverse
by a wavelength division multiplexing (WDM) metro-aggregation and often conflicting service requirements over the legacy 4G
networks. The proposed heuristic targets the minimization of 1) system will be economically unviable as dedicating network
the number of active COs (i.e, hosting DUs and/or CUs) and 2) the resources to each service will result in high capital and opera-
number of established wavelength channels under constraints of
network capacity and latency requirements. As our algorithm is
tional expenditures (CapEx and OpEx). Operators, in order to
also designed to map the RAN slices with least amount of physical provide customized network services while keeping CapEx and
resources for a given level of isolation, we also investigate the impact OpEx under control, are considering the new concept of network
of slice isolation on resource utilization. Results show how higher slicing (NS) [2], which divides a physical network into several
isolation results in higher network cost. logical virtual networks, or “slices”. Each slice is customized and
Index Terms—Flexible functional split, RAN slicing, slice dedicated to a certain type of service or to a specific application
isolation, slice mapping, WDM metro-aggregation network. in the context of a service type. On the one hand, network
slicing can help reduce CapEx (i.e., costs including equipment
investment, civil work, and installation & commissioning) by
Manuscript received July 24, 2019; revised November 3, 2019 and December sharing resources among multiple slices. On the other hand,
28, 2019; accepted February 7, 2020. Date of publication February 11, 2020; date
of current version March 17, 2020. This work was supported by the National flexible resource management enabled by software defined net-
Key R&D Program of China under Grant 2018YFB1800802, in part by the work (SDN) and network function virtualization (NFV) can also
National Nature Science Foundation of China Projects 61871051, in part by the increase the resource efficiency while reducing OpEx (i.e., costs
Beijing Natural Science Foundation under Grant 4192039, and in part by the
fund of State Key Laboratory of Advanced Optical Communication Systems including operation & maintenance (OM), energy consumption,
and Networks, China, under Grant 2019GZKF5, China Scholarship Council etc). [3].
Foundation, and in part by the European Community under Grant Agreement Nowadays, the 5G RAN architecture is still evolving to
761727 Metro-Haul project. (Corresponding author: Yuefeng Ji.)
Hao Yu, Jiawei Zhang, and Yuefeng Ji are with the State Key Labora- accommodate the concept of functional splits together with
tory of Information Photonics and Optical Communications, Beijing Uni- network slicing (5G RAN slicing). In 4G, cloud-RAN (C-RAN)
versity of Posts and Telecommunications, Beijing 100876, China (e-mail: transformed the traditional distributed RAN (D-RAN) into a
yuhao92@bupt.edu.cn; zjw@bupt.edu.cn; jyf@bupt.edu.cn).
Francesco Musumeci and Massimo Tornatore are with the Politec- 2-layer architecture by centralizing the baseband function units
nico Di Milano, Milan 20133, Italy (e-mail: francesco.musumeci@polimi.it; (BBUs) into a common pool while leaving remote radio heads
massimo.tornatore@polimi.it). (RRHs) in the cell sites (CSs). It becomes soon clear that high
Color versions of one or more of the figures in this article are available online
at http://ieeexplore.ieee.org. bandwidth demands of the fronthaul network interconnecting
Digital Object Identifier 10.1109/JLT.2020.2973311 RRHs and BBUs would make it economically unsustainable

0733-8724 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 04:30:29 UTC from IEEE Xplore. Restrictions apply.
1126 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 38, NO. 6, MARCH 15, 2020

Fig. 1. 5G RAN slicing over WDM metro-aggregation network.

to deploy C-RAN everywhere, especially in fiber-poor rural flexibility in DU/CU placement: CU and DU can be integrated
or suburban areas [4]. Hence, in 5G, new functional splits or separated. In the early stage of 5G, DU/CU integration is
have been proposed to flexibly divide RAN functions among adopted to promote low latency, low complexity operation and
centralized and distributed units depending on factors as the maintenance. However, in the mid-to-long term, the network
quality of service (QoS), latency, reliability, and network cost. will be upgraded to accommodate more challenging uRLLC
Thanks to the increased flexibility in locating functions, these and mMTC services, and DU/CU separation will become more
advanced functional splits can facilitate effective RAN slicing. beneficial [9], [10].
3GPP [5] has proposed a flexible 3-layer RAN architecture, Although network slicing can improve resource efficiency and
where three new functional blocks are created: the radio unit decrease network cost, when different tenants create slices over
(RU), distributed unit (DU) and centralized unit (CU). The the same infrastructure, inter-slice security becomes a signifi-
transport network is correspondingly divided into three parts: cant issue. Therefore, inter-slice isolation should be performed,
fronthaul, midhaul, and backhaul. where RAN functions and traffic of different slices are isolated
5G communications impose high requirements on bearer from each other for security and privacy reasons. GSMA [11] has
network bandwidth, latency, flexibility, and cost. For exam- recommended different levels of isolation among slices. Slice
ple, in 5G, peak data rate (Gbit/s) will be 20 times higher functions and network connectivity can be partly or fully isolated
and latency (ms) will be 10 times lower than the ones in 4G according to the isolation level. In contrast to network sharing,
networks [6], which imposes great challenges to transport net- isolation makes each slice operating over dedicated physical
work. For fronthaul, wavelength division multiplexing-passive resources, hence reducing resource efficiency.
optical networks (WDM-PON) is the promising solution which In this work, we focus on a new solution to perform isolation-
might be an “easy and cheap” option [7]. For midhaul and aware RAN slice mapping to improve the processing and band-
backhaul portions of the 5G transport network, optical trans- width resource utilization. As shown in Fig. 1, we consider RAN
port network (OTN) is the “best fit”, as it is already widely slicing in a multi-layer OTN over WDM metro-aggregation
deployed and can support the bandwidth, latency, and flexibility network, where several CSs, collecting mobile traffic, are dis-
requirements of 5G services. Recently, China Telecom and ZTE tributed and interconnected with central offices (COs) by optical
have cooperated on the standardization of the Mobile-optimized fiber links. COs are divided into multiple stages, depending
OTN (M-OTN) technology [8] targeting at requirements of low on their positions in the aggregation hierarchy. Specifically, at
latency, low power consumption, and low cost. In addition, the lowest network stage (say stage 0) CSs are present, while
optimal placement of DU/CU is crucial to improve resource stages 1 and 2 are composed by Access COs and Main COs,
efficiency while satisfying 5G requirements. The 3-layer ar- respectively. And a single Core CO, acting as point of presence
chitecture investigated in this paper significantly enhances the (PoP), represents the interface toward the core network segment.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 04:30:29 UTC from IEEE Xplore. Restrictions apply.
YU et al.: ISOLATION-AWARE 5G RAN SLICE MAPPING OVER WDM METRO-AGGREGATION NETWORKS 1127

These metro COs are organized in “ring-and-spur” topology

and each CO supports optical-electronic-optical (OEO) signal
conversion, which can be used to perform traffic grooming, but
introducing additional latency due to the switching and signal
conversion. Besides the switching ability, metro COs can also be
equipped with processing capacity for hosting RAN functions.
In this work, we solve the RAN function (e.g., DU, CU, and
MEC)1 placement in the metro COs, jointly considering routing
and wavelength assignment (RWA) for the traffic flows between
functions. We design a heuristic algorithm to solve the problem
with two objectives: minimizing the number of active COs and
minimizing the number of wavelengths. Then we investigate
the interaction between RAN function placement and RWA, Fig. 2. RAN architecture evolution.
and evaluate the impact of capacity/latency constraints, and
different isolation levels on the network and processing resource coordination gains, etc. Ref. [23] focused on the mapping be-
efficiency. tween radio resource blocks and optical bandwidth resource in
The rest of the paper is organized as follows. Section II the beam-forming scenario. Considering the high bandwidth re-
gives a brief overview of the evolution of RAN and flexible quirements for fronthaul, more flexible functional split between
functional split RAN architecture. In Section III, we define the RAN functions are becoming a key proposition in designing new
concept of 5G RAN and explain how slice isolation affects RAN 5G RAN architecture. Ref. [24] performed a techno-economic
slicing. In Section IV, we introduce the isolation-aware 5G RAN study on how to design a low-cost C-RAN, showing that the
slicing problem in a WDM metro-aggregation networks and proposed Hybrid-RAN with functional split can achieve up to
propose a heuristic algorithm to address this problem. Illustrative 15% cost saving compared to D-RAN. Both Refs. [25] and [26]
numerical results are shown in Section V. Section VI concludes studied baseband function placement considering different func-
the work. tional splits. Ref. [25] presented a graph-model-based functional
splitting, then investigated the deployment cost under delay
II. RELATED WORK constraints. Ref. [26] discussed different RAN functional splits
with a specific emphasis on multiplexing gain in computational
The 5G RAN architecture and the concept of network slicing
resources. To the best of our knowledge, existing works have
are still under study and standardization by several organiza-
only considered baseband function placement in 2-layer RAN,
tions, e.g., 3GPP, ITU-T, IEEE, etc. Next, we present recent
and there are no works analyzing DU/CU placement in 3-layer
related works on 5G RAN and network slicing.
RAN. Therefore, in our preliminary work [27], we modeled
a DU/CU placement problem for C-RAN deployment with a
A. RAN Evolution 3-layer RAN architecture using ILP. The results in [27] has
In recent years, C-RAN architecture has been widely inves- proved that 3-layer RAN outperforms than 2-layer RAN in the
tigated. Some studies compared D-RAN and C-RAN in terms aspect of baseband function consolidation. In this work, we
of cost and power consumption. Ref. [12], [13] emphasized the emphasize on the impact of isolation on RAN slicing in the
impact of optical communication as an enabling technology in 3-layer RAN architecture.
5G networks. Ref. [14] analyzed the total cost of ownership
(TCO) for constructing a RAN infrastructure, showing that the B. 5G Network Slice
optical infrastructure plays a dominating role in the migration Ref. [28] proposed a grooming graph for dynamic slicing
from D-RAN to C-RAN. Ref. [15] investigated CapEx when to minimize service blocking by jointly considering radio,
deploying C-RAN, showing cost saving in terms of optical transport and cloud domains. As understanding user’s behavior
switches compared to a distributed architecture. Recently, using is highly significant for effective network slicing, a big-data-
a WDM metro-aggregation network for C-RAN deployment is analytics-based slice admission policy was proposed in [29],
under investigation (see, e.g., overviews in [16] [17]), Refs. [18] where traffic-prediction-based dynamic slicing is shown to
[19] investigated the BBU placement problem over an optical achieve lower service degradation. All existing works discussed
aggregation network, and proposed ILP and heuristic solutions above investigate network slicing in traditional C-RAN or D-
for BBU placement in a 2-layer C-RAN architecture. Moreover, RAN, however, functional split and 3-layer RAN architecture
always in the context of 2-layer C-RAN, Refs. [20] [21] [22] will also have a significant impact on the performance of RAN
proposed solutions to dynamically associate RRUs and BBUs slicing and will be the subject of investigation in this work.
through reconfigurable lightpaths in fronthaul for enhancing
wireless performance metrics, such as energy efficiency, radio III. FLEXIBLE RAN ARCHITECTURE WITH FUNCTIONAL SPLIT
As shown in Fig. 2, RAN architecture is evolving towards a
1 Mobile Edge Computing server, we assume that all the services will be served more distributed functional decomposition. Following increased
in the MEC server. densification and coordination of small cells in 4G RANs, it

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 04:30:29 UTC from IEEE Xplore. Restrictions apply.
1128 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 38, NO. 6, MARCH 15, 2020

between source and destination nodes of a connection, and it is

assumed to be 5 μs/km; switching latency includes OEO and
queueing time in switching nodes. The formulas below allow to
calculate processing complexity and bandwidth requirement.

CRU = kR · B · A (1)
CDU = kD · B · A · L (2)
Fig. 3. Example of functional split options (adapted from [31].) CCU = kC · A (3)
CM EC = kM · B · L (4)

has already been proposed to split traditional BS. Centralizing D f = kF 1 · B · A · L + kF 2 · A (5)

BBUs in a pool facilitates efficient resource management and D m = kM · B · L (6)
cell coordination, which can reduce the total network cost.
Because of high bandwidth demand between BBUs and RRHs, D b = kB · B · L (7)
and emergence of diversified 5G applications, RAN is evolving where C is the processing complexity required by a given RAN
towards a more flexible architecture, a 3-layer architecture has function (DU, CU, etc.), D is the bandwidth required by the
been proposed recently [5]. Part of physical layer functions network connections, k is a coefficient related to the specific
are moved to the CS, combined with original RRH, called RAN function, B is the wireless bandwidth of the cell, and A is
RU. Higher-layer PHY functions, MAC layer and RLC layer the MIMO number of the antennas, L is the traffic load under the
constitute the DU, while the CU includes PDCP layer functions cell, which is expressed in numbers of wireless resource blocks
and user plane functions of the core network. The MEC mainly (RBs). The reference values for coefficient k for the various
includes control plane functions from the original evolved packet RAN functions are taken from [30].
core (EPC) and edge application server.
New network interfaces between RU, DU and CU are cur-
IV. ISOLATION IN RAN SLICING
rently under standardization by 3GPP [5] and other alliances [30]
[32], etc. As shown in Fig. 3, 3GPP has proposed multiple A. Definition of 5G Network Slice
functional split options, typically listed from option 1 to option A network slice represents a collection of network functions
8. Based on the proposal by 3GPP, ITU-T [33] recommends the and network connections dedicated to certain mobile services
split at Option 2 (interface F1) and Option 7 (interface Fx) as with similar requirements in bandwidth, processing, latency,
the standard split options between RU, DU and CU. reliability, etc. In this work, we assume that the mobile users
Transport bandwidth: The interface F1 at Option 2 (also within a slice are located in n adjacent CSs. For different types
known as midhaul interface) dynamically scales its bandwidth of slices, users’ traffic are quite different and can be measured
requirement with traffic load and requires only slightly more by the number of requested RBs. As an example, for eMBB
bandwidth than the backhaul interface. The bandwidth of fron- slice, mobile users need higher data rate for transmitting a large
thaul interface Fx at Option 7 also varies with the air interface amount of content, while for mMTC slice, users may not require
traffic load but requires higher rates (typically by up to an order high data rate, but need continuous data transmission during
of magnitude compared to backhaul). the whole lifecycle of slice. Similarly, for latency, an eMBB
Processing complexity: The processing complexity of base- slice requires the user plane delay not to be larger than 4 ms,
band functions can be measured in Giga Operation Per Sec- while fronthaul delay must be less than 100 μs, and midhaul
ond (GOPS) [34]. The processing complexity of low-PHY is delay must be less than 150 μs. Instead, a uRLLC slice requires
related to carrier bandwidth and the number of antennas (called stricter delay, namely, user plane delay must be less than 0.5 ms,
“cell-processing” functions, as shown in Fig. 3), whereas the while fronthaul delay must be less than 50 μs [37].
processing complexity of MAC/RLC/PDCP layer and high-
PHY layer functions are also related to the traffic load (called
“user-processing” functions). B. Isolation Requirement
Transport latency: Transport latency requirements of back- Isolation is regarded as one of the key features of network slic-
haul and midhaul links are determined by service latency re- ing. For security considerations, slice tenants may require a slice
quirements, i.e., around 10 ms for eMBB, about 1 ms for uRLLC to be fully or partly isolated from other slices. Generally, RAN
and ranging from 1 ms to several 10 ms for mMTC [35]. For slice isolation is divided into two parts: processing isolation and
the fronthaul link at Option 7, the latency is determined by the transport isolation. Processing isolation refers to isolating func-
requirements of the RAN technology. For example, a latency tions of different slices into separate physical entities (e.g., by
budget of round-trip time RT T = 3 ms is available between placing functions of different slices in different virtual machines
a DU and its corresponding RU [36] due to Hybrid Automatic (VMs) or different physical servers). Transport isolation means
Repeat reQuest (HARQ) processing. The latency contributors traffic should be transmitted in separated wavelengths, or even
modeled in this work include the propagation latency and different physical links. Different RAN isolation levels have
switching latency: propagation latency depends by the distance been defined [11], according to how many functions need to be

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 04:30:29 UTC from IEEE Xplore. Restrictions apply.
YU et al.: ISOLATION-AWARE 5G RAN SLICE MAPPING OVER WDM METRO-AGGREGATION NETWORKS 1129

latency requirement in fronthaul lf , midhaul lm and backhaul

lb , and its isolation level (to be chosen among I0 , I1 , I2 , I3 ).

V. A HEURISTIC ALGORITHM FOR ISOLATION-AWARE 5G RAN

SLICE MAPPING PROBLEM
A. Problem Statement
Since the metro-aggregation network has a hierarchical struc-
ture, it is beneficial to concentrate more network functions
into fewer COs in higher network stages to achieve higher
multiplexing gain. We define a metro CO as an “active” node
if it hosts at least one active VM (hosting DU/CU) and then
Fig. 4. Example of isolation levels. we define a metric called “consolidation ratio” R = Na /Nt ,
where Na represents the number of active nodes (COs), and Nt
TABLE I represents the number of total nodes in the networks. Lower
NETWORK REQUIREMENT OF SLICE REQUEST
values of R correspond to higher consolidation. Minimizing the
number of active nodes (and hence R) by consolidating more
DU/CU into fewer COs is beneficial to operators, especially
for energy saving. However, more consolidation (i.e., lower
R) comes at the cost of higher bandwidth consumption and
higher latency, because traffic has to traverse longer paths and
more COs. Moreover, the isolation requirements also affect
isolated, as shown in Fig. 4. For example, slice tenants might call processing/bandwidth utilization when mapping network slices
for a shared RAN with isolated core network functions, or both into metro-aggregation networks.
isolated RAN and core network functions. For isolation level Hence, in this study we investigate the isolation-aware
I0 , all functions can share physical resource among multiple 5G RAN network slice mapping problem in WDM metro-
slices, including VM resources and transport resources. In I3 , aggregation networks, that can be formally stated as follows.
all functions need to be isolated. As isolation level increases, Given i) a hierarchical multistage metro-aggregation network
more RAN functions need to be isolated. Higher isolation level topology, represented by a graph G(N, E), where N is the set
implies higher network cost because each slice is provided with of nodes (including COs and CSs) and E is the set of optical
specialized VMs and transmission bandwidth. From operators’ fiber links, and ii) slice requests defined in Section IV.A, we
perspective,2 isolation requirements can be set in two different decide how to map the RAN slice requests into a physical
ways: i) slice tenants directly ask for isolation, or ii) tenants infrastructure (mapping includes functions placement and RWA
do not express any isolation requirement, but the operator takes for fronthaul/midhaul/backhaul traffic), with two objectives: i)
decisions on isolation based on other requirements (e.g. service minimizing the number of active nodes, constrained by network
level agreements (SLA) or resource availability). We assume link capacity, latency and isolation requirement, ii) minimizing
that i) within the same operator, isolation levels are different the established wavelengths, constrained by network processing
for different slices; ii) among different operators, slices must capacity, latency and isolation requirement.
always be isolated. To achieve isolation between slices, NFV Note that, since the RU performs radio frequency and some
allows deploying network function instances to different VMs physical layer functions, it always needs to be deployed on a
on general-purpose servers, while transport isolation can be dedicated processing platform at the cell sites. Hence, placement
implemented by OTN/WDM technology and FlexE technology is only devised for DU, CU, and MEC.
which can provide independent wavelengths or time slots for
different connections. B. Algorithm Description
In summary, according to the requirements on processing
complexity, transport bandwidth,3 latency, and isolation level, Fig. 5 illustrates an example of RAN slice mapping. We
the properties of 5G RAN slice are summarized in Table I. assume mobile users are of two types, uRLLC users and eMBB
Each generic slice request s is defined by its slice coverage, users, which belong to different operators (colored by red,
represented by a set of n cells (c1 , c2 , . . . , cn ), the number yellow, blue and green). Fig. 5(a) shows how the processing
of RBs requested per cell, represented by (r1 , r2 , . . . , rn ), its isolation works when mapping RAN functions. Due to low
latency of a uRLLC service, all its functions will be placed
in COs near the cell sites, see, e.g., the placement of blue and
2 In this work, we assume that there is only one physical infrastructure green users in Fig. 5(a). For example, note that, in the case of the
owned by infrastructure provider (InP), several operators buy/lease the physical DUs for blue users, due to latency constraints, all the functions
resource and rent the resource to the slice tenants. can only be placed in access CO 1. For the eMBB users, the
3 We use the requested RBs to represent the processing complexity and
transport bandwidth, and refer to [30] for how to calculate processing complexity functions can be deployed at higher-stage COs to achieve higher
and transport bandwidth according to the number of RBs. processing consolidation as latency constraint is looser. Placing

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 04:30:29 UTC from IEEE Xplore. Restrictions apply.
1130 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 38, NO. 6, MARCH 15, 2020

Fig. 5. Illustration of 5G RAN slice mapping.

functions in core CO enables higher DU/CU consolidation, but problem [38], hence, to find an scalable solution, we divide it
also higher transport latency. Concerning isolation, the blue DU into two objectives: 1) minimize the number of active nodes and
and yellow DU are located in the same VM (dashed square) in 2) minimize the established wavelengths in the network. When
access CO 1, however, the green DU is located in another VM, considering objective 1, we set bandwidth capacity in each link
because green users belong to the different operators from blue as limited and the computational resource as always enough,
and yellow users. Fig. 5(b) shows the illustration of bandwidth so we can target at the minimization of active nodes. On the
isolation which also implies the relationship between DU/CU contrary, when considering objective 2, we set the computational
consolidation and bandwidth consumption, more consolidation resource in each CO as limited and the bandwidth resource as
comes at the cost of more bandwidth consumption. always enough, so we can aim at the minimization of established
The proposed algorithm addresses the 5G RAN slice mapping wavelengths in the network.
problem in a static scenario, i.e., when all the slice requests are The execution of the algorithm is detailed in Algorithm 1.
given, the algorithm aims at accommodating all the slice requests The characteristics of slice requests are given by the parameters
and targets at the minimization of network and computational shown in Table I, and the network parameters are summarized
resource. According to problem statement, the network slicing in Table II.
problem in this work can be regarded as a service function 1) Node Ranking: For function mapping, the problem is
chain (SFC) mapping problem which can be divided into node solved by using existing bin-packing methodologies. All the
mapping and link mapping. As for node mapping, we solve it physical nodes are ranked in a certain order considering the
using bin-packing methodologies. Then after determining the residual processing resource of nodes, the residual bandwidth
locations of RAN functions, the connections between functions of optical ports of nodes, the distance to the source node, and
need to be mapped. This problem is a routing and wavelength the node level (Access CO, Main CO, and Core CO) in the
assignment problem in a multilayer OTN over WDM aggre- metro-aggregation network, then the RAN functions are placed
gation network. As we know, SFC problem is an NP-Hard into the physical nodes with highest ranking value. To this end,

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 04:30:29 UTC from IEEE Xplore. Restrictions apply.
YU et al.: ISOLATION-AWARE 5G RAN SLICE MAPPING OVER WDM METRO-AGGREGATION NETWORKS 1131

TABLE II
Algorithm 1: 5G RAN Slice Mapping Scheme. NETWORK PARAMETERS
Input: slice requests S, physical topology G(N, E)
Output: RAN functions placement, routing and
wavelength allocation of each slice
1: Sort the slice requests in S in descending order
according to the total amount of RBs requested by
each slice, put the sorted slice requests into S 
2: Set the parameter α, β, λ, γ according to the type of
slice request, calculate the node evaluation factor Fn ,
put the ranked nodes into N 
3: for all slice request s in S  do ,
4: sort the CSs belonging to s in descending order 33: if the slice s in VM v and slice s belong to the
according to the requested RBs, put the sorted CSs same operator and the isolation level of slice s is also
into C  I0 ∼ I2 then
5: for all CS c in C  do 34: map the function into this VM v
6: calculate the processing requirement of RU, DU, 35: else
CU and MEC function and bandwidth requirement 36: create a new VM to instantiate this function
of fronthaul, midhaul and backhaul with formula 37: end if
(1) ∼ (7). For each CS, RU function is deployed 38: else
locally in CS. 39: create a new VM to instantiate this function
7: for each function f in DU, CU, MEC do 40: end if
8: Objective 1: Minimize the active nodes 41: end Function
9: for all node n in N  do 42: Function MapLink(lightpath p, physical links l)
10: if (1) there is at least one lightpath between 43: for all candidate links n in the path L do
node n and CS c whose latency is lower than 44: if the isolation level is I0 ∼ I2 then
lf /lm /lb , (2) if the lightpath exists, whether there is 45: find a existing active wavelength w with enough
enough bandwidth available in each link on this bandwidth capacity
ligthtpath l then 46: if the slice s in wavelength w and slice s belong
11: MapFunction(f , n ) to the same operator and the isolation level of slice s
12: MapLink(l, physical links) is also I0 ∼ I2 then
13: end if 47: map the lightpath into this wavelength w
14: end for 48: else
15: Objective 2: Minimize the active wavelengths 49: create a new wavelength to map this lightpath
16: for all node n in N  do 50: end if
17: if (1) whether there is at least one lightpath 51: else
between node n and CS c whose latency is lower than 52: create a new wavelength to map this lightpath
lf /lm /lb , (2) if there is enough processing capacity in 53: end if
the node n then 54: end for
18: put node n into set N  55: end Function
19: end if
20: for each node n in N  do
21: calculate the number of links between node n and CS
c and choose the lightpath l with minimum number of we propose the following node ranking method:
links between node n∗ and CS c
22: end for Fn = αUn + βBn + λHn + γLn (8)
23: MapFunction(f , n∗ )
24: MapLink(l, physical links) where Fn is the comprehensive evaluation index of the physical
25: end for node n, Un is the used processing resource of node n, and Bn
the used bandwidth resource by all the ports of node n, Bn =
26: end for is
p∈P (n) BW (p), BW (p) is used bandwidth of each port of
27: end for
28: end for node n. Hn is the average distance between  the node n and
29: Function MapFunction(function f, physical node n) dis(r)
all the source node in slice request, Hn = r∈R|R| , dis(r)
30: check the isolation level of this slice request s is the shortest distance between node n and each cell site r. R
31: if the isolation level is I0 ∼ I2 then is the set of source nodes in slice request. Ln is the stage level
32: find a existing active VM v with enough processing of node n in the metro networks. α, β, λ, γ are coefficients and
capacity α + β + λ + γ = 1.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 04:30:29 UTC from IEEE Xplore. Restrictions apply.
1132 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 38, NO. 6, MARCH 15, 2020

The coefficient α, β, λ, γ can be adjusted according to dif-

ferent slice types and algorithm objectives. For example, for
the eMBB slice, the value of α can be larger, so that the nodes
with more remaining processing resource are more likely to be
chosen. For uRLLC type slice, the value of β should be larger
compared to other coefficients, so that the nodes closer to the
cell sites are more likely to be selected.
2) RAN slice mapping: Firstly, the algorithm sorts the slice
requests in S in descending order according to the total amount
of RBs requested by each slice, put the sorted slice requests into
S  , then iterates the slice requests in the set S  (line 1). Next, we
perform node ranking (line 2) for current slice request s, then
sort the CSs of slice s in descending order of requested RBs in
each CS, then put the sorted CSs into set C  (line 4). Next, we
pop up the CS c in C  , calculate the processing requirement of
DU, CU and MEC and the bandwidth requirement of fronthaul,
midhaul and backhaul using formula (1) ∼ (7) (line 6).
r Objective 1: Minimize the active nodes. For the function Fig. 6. Large-scale network topology.
f in {DU, CU, M EC}, check whether there is at least one
path between node n in set N  and CS c whose distance
is lower than lf /lm /lb , if there exists such a path, check which satisfy the constraints, calculate the number of
if there is enough available bandwidth in each link along links of the path between candidate nodes and CS c,
the path (line 10). Then try to place function f in the choose the node which has a minimum number of links
node n with M apF unction(f, n ) (line 11), otherwise, to CS c in order to avoid establishing more wavelengths
iterate next node in N  . During the placement of function along the lightpath. Once the candidate node is decided,
f into the physical nodes, isolation is the main considera- call the M apF unction(f, n ) and M apLink(l, n∗ ) to
tion when performing M apF unction(f, n ). First, we try finish the node mapping and link mapping.
to re-use the existing active VM v, since reusing active
VMs can maximize utilization of processing resources, VI. ILLUSTRATIVE NUMERICAL RESULTS
then we check if there is an isolation conflict between
the function f and the functions already in the VM v. (1) A. Case Study
Check if the function f belongs to the same operator with We initially consider a small-scale RAN slice deployment
functions in VM v. (2) According to the function types, scenario, with 9 metro nodes and 6 cell sites in the network, at
check the isolation level. For DU, if the already placed most 2 wavelengths at 1Gbps in each fiber link and VM capacity
functions in v and function f are both with isolation level set to 100 GOPS. The radio configuration of each sector of a cell
I0 ∼ I2 , then function f can be placed in VM v sharing site is 20 MHz carrier bandwidth, QPSK, and 2 × 2 MIMO and
the processing resource with other slices. If either function 100 RB under this configuration (1 RB for 180 kHz and 7 OFDM
is with isolation level I3 or their operators are different, symbols). All the slice requests belong to the same operator.
the function f should be placed by creating a new VM Each slice request contains 5 cell sites, and the requested RBs
(line 29–41). Link mapping is after the function f being follow a normal distribution with an average 10 and deviation 5.
placed, since there is already an available path satisfying In addition, there is no isolation requirement in this case. Using
the latency and bandwidth constraints, then the isolation is this simplified case, we validate the performance of our heuristic
the only consideration when mapping connection l between against the ILP in [27].
function f and its former function into physical links. Then we consider the large-scale topology in Fig. 6, with 52
Just like function placement, transport isolation means two metro nodes, consisting of 2 core COs, 6 main COs, 44 access
connections that need to be isolated should be mapped into COs, each access CO is associated with 3 CSs, thus 132 CSs in
two separated wavelengths. For fronthaul, if the connec- this topology (not shown in the figure for the sake of clarity).
tion l and connections in existing wavelengths are both At most 40 wavelengths at 10Gbps can be established in each
with isolation level I0 ∼ I2 , then map this connection l, fiber link. VM capacity is 1000 GOPS and node capacity is
otherwise, establish a new wavelength. 3/10/20/50 VMs for cell sites/access CO/main CO/core CO. The
r Objective 2: Minimize the established wavelengths. radio configuration is in line with the guidelines of a 5G urban
Different from objective 1, the minimization of estab- aggregation network, as specified in [39], which is 100 MHz,
lished wavelengths requires another solution for link map- 256 QAM and 8 × 8 MIMO. Under this radio configuration,
ping. In objective 2, the capacity of physical nodes is 500 RBs are available in each cell. Multiple slices can share
set to be limited and bandwidth capacity is set infinite. these 500 RBs in each CS, which means that the number of
In the node mapping phase, after node ranking, all the requested RBs of all the slices in one CS can not exceed 500.
ranked nodes are checked if satisfying with processing We assume that the requested RBs of each CS within one slice
capacity and latency constraints. Then, among the nodes are normally distributed, for the eMBB/uRLLC/mMTC type,

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 04:30:29 UTC from IEEE Xplore. Restrictions apply.
YU et al.: ISOLATION-AWARE 5G RAN SLICE MAPPING OVER WDM METRO-AGGREGATION NETWORKS 1133

Fig. 7. (a) ILP vs. heuristic in R with a small topology; (b) R values vs. maximum service latency requirement (left: 2-layer, right: 3-layer); (c) R value vs.
maximum service latency requirement with different W .

TABLE III (3) number of established wavelengths, and we study the impact
LATENCY REQUIREMENT OF FRONT/MID/BACKHAUL
FOR EMBB/URLLC/MMTC TYPE
of slice demand, slice types, isolation level and network capacity
on these metrics.
1) Small Topology: We solve the RAN slice mapping prob-
lem using ILP and heuristic with small topology. The details of
ILP are shown in our previous work [27] and the heuristic only
performs objective 1. In this case, no isolation is performed and
no specific slice type is enforced. Fig. 7 shows the comparison
TABLE IV
REFERENCE VALUE OF α, β, λ, γ FOR DIFFERENT OBJECTIVES AND DIFFERENT
between ILP and heuristic in terms of consolidation factor R.
SLICE TYPES Lower value for R represents more consolidation, i.e., less COs
are activated to host the RAN functions. As shown in Fig. 7(a),
ILP achieves the best performance (smaller values) in terms of
R, but heuristic results lay approximatively within 10% of the
ILP. Leveraging its finer granularity in placing RAN function,
3-layer architecture achieves a better performance than in 2-layer
architecture. In fact, when placing functions, the functions with
less processing requirements are more likely to be mapped into
the VMs without exceeding residual capacity.
the average number of requested RBs is 100/30/30. Moreover, Besides residual capacity, the RAN function placement also
the number of CSs in a slice is uniformly distributed between depends on other factors, e.g., latency requirement (in this case,
[5, 10]/[5, 10]/[20, 30] for eMBB/uRLLC/mMTC type. Three we set the user plane latency requirement between RU and MEC
operators are considered in this case. The maximum tolerated varying from 100 μs to 1000 μs, and the fronthaul latency
fronthaul/midhaul/backhaul latency for eMBB/uRLLC/mMTC requirement is fixed value with 250 μs [42] in this case). As
type under the considered RAN split are shown in the Table III, shown in Fig. 7(b), the value of R increases and reaches a maxi-
derived from [40], [41]. mum value at 500 us, then decreases with the maximum latency
Node ranking is essential for the heuristic since it provides a requirement. When the latency requirement is strict, almost all
criterion during the node mapping phase, which will affect the the functions are placed in the cell sites (red bar). When the
performance of the heuristic. Therefore, the parameters in (8) latency requirement is relatively loose, part of functions (e.g.,
should be set accordingly.4 The value of parameter α, β, λ, γ CU and MEC) tend to be placed in the higher stage of networks,
for different slice types is presented in the Table IV. All the while few functions (e.g., DU) are still located in cell site due
parameters used in the numerical evaluation are summarized in to the strict fronthual latency requirement, thus the total active
Table V. COs increase. When the latency requirement is loose enough,
most of the functions will be placed in the core COs remaining
B. Discussion few DUs in the cell sites.
In the following numerical evaluation, we define the following Fig. 7(c) shows the impact of the number of wavelengths
metrics: (1) consolidation factor R; (2) number of created VMs; on the consolidation factor R. Besides the latency requirement,
the link bandwidth capacity also influences the RAN function
4 For the objective 1, minimization of active nodes is the main objective,
placement. For instance, fronthaul connections are characterized
the weight of node level should be set highest and weight of used processing by high bandwidth demand, shortage of bandwidth capacity in
resource is set to be 0, other parameters could be set according to the slice physical links will force the RAN functions to be placed in the
types. For example, the weight of bandwidth is supposed to be higher in order COs near the CSs. Therefore, the more bandwidth capacity in
to increase the possibility of multiplexing gain in bandwidth for eMBB case,
the weight of latency should be set higher in order to increase the possibility of physical links, the more consolidation in processing resource.
mapping uRLLC slice. In addition, 3-layer architecture is also proved with a better

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 04:30:29 UTC from IEEE Xplore. Restrictions apply.
1134 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 38, NO. 6, MARCH 15, 2020

TABLE V
PARAMETERS USED IN THE NUMERICAL EVALUATION

Fig. 8. (a) R value with increasing slice request for different slice types; (b) R value with increasing slice request with different W .

performance in function consolidation because of the flexible R in Fig. 8(b). If W = 1, the number of active nodes increases
RAN functions placement and traffic routing than 2-layer archi- rapidly with slice requests, when the number of slice requests is
tecture. 100, almost all the CSs and part of COs are active for processing
2) Large Topology: In this case, we solve the problem only the RAN functions since there is not enough transport band-
by a heuristic as the ILP cannot solve this case in a reasonable width, most of the functions must be placed locally. If W  10,
time. In Fig. 8(a), we compare the performance of heuristic the curves rise smoothly and more processing consolidation can
in terms of R for 3 slice types: eMBB, uRLLC and mMTC. be achieved.
It shows that uRLLC slices activate most COs, which is the Then, the impact of slice isolation as a key requirement when
lowest consolidation in processing resource since the latency mapping RAN slices is evaluated, and we also investigate the
requirements of uRLLC slices are much stricter than the other computational resource consolidation under VM isolation. To
slice types. As the latency requirement of mMTC slices is the clearly show the impact of isolation on resource utilization, we
loosest among the three types, mMTC case achieves the highest change the metric from R to the number of created VMs. We
resource consolidation. Note that, to clearly show the impact investigate the number of created VMs with the increasing slice
of latency requirement on processing resource consolidation, requests for different types of slices. As shown in Fig. 9(a), due
we set the overlay as a metro-aggregation network transport to the huge traffic demand of eMBB service, the number of VMs
solution, where each traffic flow is allocated with dedicated created for eMBB slices is the highest among the three types.
wavelength without traffic grooming. Therefore, the slice type On the other hand, we can observe that the better performance
(also latency requirement) is the only impacting factor in this of 3-layer architecture measured in the number of active VMs.
case. Especially, processing consolidation of eMBB case with 3-layer
Next, the results show the impact of transport network capac- architecture is the highest among three slice types, e.g., almost
ity (i.e., number of wavelength W ) on the consolidation factor 28% less created VMs for 1200 slice requests.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 04:30:29 UTC from IEEE Xplore. Restrictions apply.
YU et al.: ISOLATION-AWARE 5G RAN SLICE MAPPING OVER WDM METRO-AGGREGATION NETWORKS 1135

Fig. 9. (a) Number of VMs vs. slice requests for different slice types; (b) Number of VMs vs. slice requests with different isolation levels; (c) Number of VMs
vs. number of operators for different slice types.

Fig. 10. Number of wavelengths vs. slice requests for different slice types.
Fig. 11. Number of wavelengths vs. slice requests with different isolation
levels.

In the results above, the isolation level of all the slice requests
is set to be I0 , which means no isolation is adopted, all the RAN 2-layer and 3-layer seems not as obvious as in created VMs.
functions share the whole physical infrastructure as long as there Secondly, the number of wavelengths of different slice types
are enough available capacity in the VMs. In the following, the shows different trends at different stages. For eMBB slices,
number of created VMs with different isolation levels are shown. because of the high bandwidth demand, a large amount of
In this case, we consider three operators and three slice types wavelengths are established in the beginning stage. However,
mixed randomly. As shown in Fig. 9(b), when the isolation level for uRLLC slices, the established wavelengths increase rapidly
changes from I0 to I3 , the number of created VMs increases and exceed the ones for eMBB slices when the traffic load is
by 6 times on I3 with respect to I0 . In addition, we set the high. As we know, different slice types mean different latency
number of slice requests and isolation level fixed to show the requirements, because of the low latency property, the RAN
impact of the number of operators on created VMs for both functions of uRLLC slices are more likely to be distributed in
uRLLC and eMBB, as shown in Fig. 9(c). We find that the the networks, there will be less possibility for the multiplexing
number of operators has not a prominent influence as much as of bandwidth in fewer wavelengths. On the contrary, more
the isolation level on the created VMs. We can envision that connections of eMBB slices can be multiplexed in the higher
more operators (possibly, virtual operators) can co-exist in the stages of metro networks. Although in the beginning, eMBB
same infrastructure sharing the whole network resource without slices consume more wavelengths because of high bandwidth
excessive cost increase, even if all the virtual operators operate demand, the uRLLC slice will consume many more wavelengths
separately. than the eMBB case eventually.
In Fig. 10, we perform the objective 2 to minimize the Finally, we present the number of established wavelengths
established wavelengths (the sum of active wavelengths in all under different isolation levels. In this case, three types of slices
the physical links) and evaluate the relationship between the are randomly mixed. As shown in Fig. 11, the result shows that
number of established wavelengths and slice types. In this case, higher isolation level, more wavelengths established. Note that
we set the isolation levels of slices as level I3 . Firstly, we can the type of isolation level does not have a significant effect
observe that the difference in established wavelengths between on the number of wavelengths (in fact, the four curves are

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 04:30:29 UTC from IEEE Xplore. Restrictions apply.
1136 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 38, NO. 6, MARCH 15, 2020

very close to each other) at low traffic load. Because a large [15] L. Velasco et al., “Meeting the requirements to deploy cloud RAN over
number of new wavelengths are established in the beginning, optical networks,” IEEE/OSA J. Opt. Commun. Netw., vol. 9, no. 3,
pp. B22-B32, Mar. 2017.
with the increasing of slice requests, the slice tends to be mapped [16] T. Pfeiffer, “Next generation mobile fronthaul and midhaul architectures,”
in existing wavelengths, so the curves increase slowly after a IEEE/OSA J. Opt. Commun. Netw., vol. 7, no. 11, pp. B38–B45, Nov. 2015.
certain point. [17] X. Liu and F. Effenberger, “Emerging optical access network technologies
for 5G wireless,” IEEE/OSA J. Opt. Commun. Netw., vol. 8, no. 12,
pp. B70–B79, Dec. 2016.
VII. CONCLUSION [18] F. Musumeci, C. Bellanzon, N. Carapellese, M. Tornatore, A. Pattavina,
and S. Gosselin, “Optimal BBU placement for 5G C-RAN deployment
In this paper, we investigate the problem of 5G RAN slicing over WDM aggregation networks,” J. Lightw. Technol., vol. 34, no. 8,
considering a functional-split based 3-layer RAN architecture pp. 1963–1970, Apr. 2016.
[19] F. Musumeci, O. Ayoub, M. Magoni, and M. Tornatore, “Latency-aware
over WDM metro-aggregation networks, with dual-objective CU placement/handover in dynamic WDM access-aggregation networks,”
of 1) minimizing the number of active metro nodes, 2) min- in IEEE/OSA J. Opt. Commun. and Netw., vol. 11, no. 4, pp. B71–B82,
imizing the number of established wavelengths. To solve this Apr. 2019.
[20] J. Zhang, Y. Ji, X. Xu, H. Li, Y. Zhao, and J. Zhang, “Energy efficient
problem, we propose a RAN slice mapping heuristic for the RAN baseband unit aggregation in cloud radio and optical access networks,”
functions placement and RWA of mobile traffic considering IEEE/OSA J. Opt. Commun. Netw., vol. 8, no. 11, pp. 893–901, Nov. 2016.
the capacity and latency constraints, as well as slice isolation [21] J. Zhang, Y. Ji, S. Jia, H. Li, X. Yu, and X. Wang, “Reconfigurable optical
mobile fronthaul networks for coordinated multipoint transmission and
requirements. Results show that the RAN slice mapping in reception in 5G,” IEEE/OSA J. Opt. Commun. Netw., vol. 9, no. 6, pp. 489–
3-layer architecture outperforms the one in 2-layer architecture 497, Jun. 2017.
in terms of processing resource consolidation (e.g., created VMs [22] H. Yu, J. Zhang, Y. Ji, and M. Tornatore, “Energy-efficient dynamic light-
path adjustment in a decomposed AWGR-based passive WDM fronthaul,”
decrease by 28% less with 3-layer architecture on eMBB case), IEEE/OSA J. Opt. Commun. Netw., vol. 10, no. 9, pp. 749–759, Sep. 2018.
due to the more flexible placement with a finer granularity of [23] J. Zhang, Y. Xiao, D. Song, L. Bai, and Y. Ji, “Joint wavelength, an-
functions in 3-layer architecture. Finally, the trade-off between tenna, and radio resource block allocation for massive MIMO enabled
beamforming in a TWDM-PON based fronthaul,” IEEE/OSA J. Lightw.
network sharing and slice isolation is investigated, and the results Technol., vol. 37, no. 4, pp. 1396–1407, Feb. 2019.
show that the slice isolation is achieved at the cost of higher [24] X. Wang, L. Wang, S. E. Elayoubi, A. Conte, B. Mukherjee, and C. Cavdar,
resource occupation (6 times more VMs on I3 case than I0 case). “Centralize or distribute? A techno-economic study to design a low-cost
cloud radio access network,” in Proc. IEEE Int. Conf. Commun., 2017,
pp. 1–7.
[25] J. Liu, S. Zhou, J. Gong, Z. Niu, and S. Xu, “Graph-based framework for
REFERENCES flexible baseband function splitting and placement in in C-RAN,” Proc.
[1] NGMN, “NGMN 5G White Paper,” NGMN Alliance, Frankfurt am Main, IEEE Int. Conf. Commun., 2015, pp. 1958–1963.
Germany, 2015. [26] M. Shehata, A. Elbanna, F. Musumeci, and M. Tornatore, “Multiplexing
[2] NGMN, “Description of Network Slicing Concept,” NGMN Alliance, gain and processing savings of 5G radio-access-network functional splits,”
Frankfurt am Main, Germany, Jan. 2016. IEEE Trans. Green Commun. Netw., vol. 2, no. 4, pp. 982–991, Dec. 2018.
[3] M. R. Raza, M. Fiorani, A. Rostami, P. Öhlen, L. Wosinska and P. Monti, [27] Y. ji, H. Yu, J. Zhang, Y. Xiao, F. Musumeci, and M. Tornatore, “DU/CU
“Demonstration of dynamic resource sharing benefits in an optical C- Placement for C-RAN over optical metro-aggregation networks,” in Proc.
RAN,” IEEE/OSA J. Opt. Commun. Netw., vol. 8, no. 8, pp. 621–632, Conf. Opt. Netw. Des. Modelling (ONDM), 2019, pp. 1–6.
Aug. 2016. [28] M. R. Raza, M. Fiorani, A. Rostami, P. Öhlen, L. Wosinska, and
[4] “Assessment of candidate transport network architectures for structural P. Monti, “Dynamic slicing approach for multi-tenant 5G transport net-
convergence” Convergence of fixed and Mobile Broadband access/ ag- works [invited],” IEEE/OSA J. Opt. Commun. Netw., vol. 10, no. 1,
gregation networks—COMBO Project, Jun. 2016. [Online]. Available: pp. A77–A90, Jan. 2018.
https://www.ict-combo.eu [29] M. R. Raza, A. Rostami, L. Wosinska, and P. Monti, “A slice admission
[5] “Radio access architecture and interfaces,” (Release 14), 3GPP, Sophia policy based on big data analytics for multi-tenant 5G networks,” J. Lightw.
Antipolis Cedex, France, Tech Rep. TR 38.801 V14.0.0, 2017. Technol., vol. 37, no. 7, pp. 1690–1697, Apr. 2019.
[6] Series, M, “IMT Vision—Framework and overall objectives of the future [30] “Small Cell Virtualization: Functional Splits and Use Cases,” Small Cell
development of IMT for 2020 and beyond,” Recommendation ITU Vision: Forum White Paper, Rel. 7, Jul. 2016.
2083-0, 2015. [31] “Transport network support of IMT-2020/5G,,” ITU-T Study Group 15
[7] N. J. Gomes, P. Chanclou, P. Turnbull, A. Magee, and V. Jungnickel, meeting, Geneva, Tech. Rep. ITU-T GSTR-TN5G, Jan. 29, 2018.
“Fronthaul evolution: from CPRI to Ethernet,” Opt. Fiber Technol., vol. 26, [32] Next generation Mobile Network (NGMN) Alliance, “Project RAN evo-
pp, 50–58, 2015. lution: Further study on critical C-RAN technologies,” Mar. 2015.
[8] “China Telecom’s Requirements on 5G Transport,” China Telecom, [33] G.sup.5GP, “5G Wireless Fronthaul Requirements in a PON Context,”
2017. [Online] Available: https://www.itu.int/en/ITU-T/Workshops-and- ITU-T, G.Sup66 Supplement, Jul. 2019. [Online]. Available: https://www.
Seminars/20171016/Documents/3.Rick.pdf itu.int/rec/T-REC-G.Sup66-201907-I/en
[9] “China Telecom 5G Technology White Paper,” June 2018. [Online] [34] B. Debaillie, C. Desset, and F. Louagie, “A flexible and future-proof power
Available: http://www.chinatelecom.com.cn/2018/ct5g/201806/ model for cellular base stations,” in Proc. IEEE 81st Veh. Technol. Conf.,
P020180626325685163826.pdf Glasgow, Scotland, 2015, pp. 1–7.
[10] “The transport network considerations for 5G in CMCC,” 2018. [On- [35] Nokia, “5G new radio network: Use cases, spectrum, technologies
line]. Available: https://www.opennetworking.org/wp-content/uploads/ and architecture,” 2019. [Online]. Available: https://www.scribd.com/
2018/12/The-transport-network-consi-deration-for-5G-in-CMCC.pdf document/402285516/Nokia-5G-New-Radio-white-paper-EN-PDF
[11] “Network Slicing Use Case Requirements Whitepaper,” GSMA, London, [36] 3GPP TS-36.213 (Physical layer procedures). (2015). [Online]. Available:
U.K., Apr. 2018. http://www.3gpp.org
[12] Y. Ji, J. Zhang, X. Wang and H. Yu, “Towards converged, collaborative and [37] “New-Transport-Network-Architectures-for-5G-RAN,” Fujitsu, Tokyo,
co-automatic (3C) optical networks,” Sci. China Inf. Sci., vol. 61, no. 12, Japan, 2018.
pp. 11301, 2018. [38] S. Khebbache, M. Hadji, and D. Zeghlache, “Virtualized network functions
[13] Y. Ji, J. Zhang, Y. Xiao and Z. Liu, “5G flexible optical transport networks chaining and routing algorithms,” Comput. Netw., vol. 114, pp. 95–110,
with large-capacity, low-latency and high-efficiency,” China Commun., 2017.
vol. 16, no. 5, pp. 19–32, May 2019. [39] “Deliverable 3.3: Analysis of transport network architectures for struc-
[14] S. S. Lisi, A. Alabbasi, M. Tornatore and C. Cavdar, “Cost-effective tural convergence,” Convergence of fixed and Mobile Broadband access/
migration towards C-RAN with optimal fronthaul design,” in Proc. IEEE aggregation networks—COMBO Project, Jul. 2015. [Online]. Available:
Int. Conf. Commun., Paris, 2017, pp. 1–7. https://www.ict-combo.eu

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 04:30:29 UTC from IEEE Xplore. Restrictions apply.
YU et al.: ISOLATION-AWARE 5G RAN SLICE MAPPING OVER WDM METRO-AGGREGATION NETWORKS 1137

[40] F. Alam, “Dimensioning Challenges of xhaul,” IEEE P1914.1 TF Massimo Tornatore (Senior Member, IEEE) is an Associate Professor with
reference, Mar. 20, 2018. [Online]. Available: http://sagroups.ieee. the Department of Electronics, Information, and Bioengineering, Politecnico di
org/1914/wp-content/uploads/sites/92/2018/03/tf1_1803_Alam_xhaul- Milano, Milano, Italy. He also holds an appointment as an adjunct Associate
dimensioning-challenges_1.pdf. Professor with the Department of Computer Science, University of California,
[41] “Transport requirement for CU&DU functional splits options,” R3- Davis and the D. R. Cheriton School of Computer Science, University of Wa-
161813, 3GPP TSG RAN WG3 Meeting #93, Göteborg, Sweden, terloo. He is an Author of more than 300 peer-reviewed conference and journal
Aug. 22, 2016. [Online]. Available: https://portal.3gpp.org/ngppapp/ papers. His research interests include performance evaluation, optimization and
CreateTDoc.aspx?mode=view&contributionUid=R3-162005 design of communication networks (with an emphasis on the application of
[42] H. Son and S. Shin, “Fronthaul Size: Calculation of maximum dis- optical networking technologies), cloud computing, and automated network
tance between RRH and BBU,” NETMANIAS. 2014. [Online]. Avail- management using machine learning. He is a member of the editorial board
able: https://www.netmanias.com/en/post/blog/6276/c-ran-fronthaul-lte/ of Photonic Network Communications, Optical Switching and Networking,
fronthaul-size-calculation-of-maximum-distance-between-rrh-and-bbu IEEE COMMUNICATION LETTERS and IEEE COMMUNICATION SURVEYS & TU-
TORIALS. He is an active member of the Technical Program Committees of
various networking conferences such as IEEE INFOCOM, OFC, ICC, and
Hao Yu (Student Member, IEEE) received the B.S. degree in communication GLOBECOM. He has been a co-recipient of 17 Best Paper awards.
engineering from the Beijing University of Posts and Telecommunications
(BUPT), Beijing, China, in 2015. He is currently working toward the Ph.D.
degree in information and communication engineering with BUPT, under the
supervision of Prof. Yuefeng Ji. He was a Joint-Supervised Ph.D. Student with
the Politecnico di Milano, Milano, Italy. His research interests include 5G optical
transport networks, software defined optical network, intelligent optical network.

Francesco Musumeci (Member, IEEE) received the Ph.D. degree in information

engineering from Politecnico di Milano, Milano, Italy, in 2013, where he is
currently an Assistant Professor with the Department of Electronics, Information
and Bioengineering. He is the Author of more than 70 papers in the area of Yuefeng Ji (Senior Member, IEEE) received the Ph.D. degree from the Beijing
communication networks, published in international journals and conference University of Posts and Telecommunications, Beijing, China. He is currently
proceedings, and is co-winner of three Best Paper Awards from IEEE sponsored a Professor and the Deputy Director of the State Key Lab of Information
conferences. His current research interests include design, optimization and Photonics and Optical Communications. His research interests are primarily
performance analysis of optical networks, applications of machine learning to in the area of broadband communication networks and optical communications,
networking, 5G, network function virtualization and network disaster resilience. with emphasis on key theory, realization of technology, and applications. He is
a Fellow of China Institute of Communications.

Jiawei Zhang received the Ph.D. degree from the State Key Laboratory of
Information Photonics and Optical Communications, Beijing University of
Posts and Telecommunications (BUPT), Beijing, China. He is currently an
Associate Professor with BUPT. He was a Joint-Supervised Ph.D. Student with
the University of California at Davis, Davis, CA, USA. His current research
interests include 5G RAN transport networks, network function virtualization,
software-defined radio, and optical access networks, with an emphasis on the var-
ious fronthaul-/midhaul-/backhaul-based optical communication technologies.
He served on the Technical Program Committees for the IEEE DRCN 2019 and
for the Workshop on Cloud Computing Systems, Networks, and Applications
at the IEEE GLOBECOM 2014, ICC 2015, and INFOCOM 2018 conferences.
He also severed as a Guest Editor of the Special Issue Resilience in future 5G
Photonic Networks of Photonic Network Communications Journal (Springer).

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 04:30:29 UTC from IEEE Xplore. Restrictions apply.