Toward End To End Latency Management of 5G Network Slicin 2023 Optical Fiber

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

Optical Fiber Technology 76 (2023) 103220

Contents lists available at ScienceDirect

Optical Fiber Technology


journal homepage: www.elsevier.com/locate/yofte

Regular Articles

Toward end-to-end latency management of 5G network slicing and fronthaul


traffic (Invited paper)
David Larrabeiti a ,∗, Luis M. Contreras b , Gabriel Otero a , José Alberto Hernández a ,
Juan P. Fernandez-Palacios b
a
Universidad Carlos III de Madrid, Av. Universidad 30, Leganes, E-28911, Madrid, Spain
b Telefonica I+D/CTIO, Madrid, Spain

ARTICLE INFO ABSTRACT

Keywords: 5G network slicing allows operators to deploy virtual connectivity services tailored for specific purposes on
5G slicing top of the same underlying physical infrastructure. For some 5G services, the telecommunication operator
Latency management needs to provide the customer with real-time information of the end-to-end Quality of Service for a particular
Fronthaul transport
slice. This paper focuses on the challenges behind this target, how per-layer per-segment monitoring can be
Backhaul
performed based on common open interfaces to standard OAM protocols, and provides practical rules to plan
Tactile services
end-to-end latency for slices. Then it reviews a few latency engineering approaches for fronthaul traffic from
work carried out in this research area.

1. Introduction users, corporate workers, roamers, or another public mobile network


operator). Furthermore, 5G features network capability exposure to
3GPP has identified network slicing as a key component of 5G. Tech- trusted third parties through an appropriate API. That API shall allow
nical specification TS23.501, System architecture for the 5G System to create, modify or delete slices, associate UEs to slices, and also im-
(5GS), includes slicing since 5G release 16 at stage 2.1 TS22.261 [1] for portantly, to monitor the network slice, including resource utilization
release 18, also at stage 2, specifies the support of slicing as one of the and QoS. A recognized target slice by 3GPP is the implementation of
formal service and operational requirements for 5G systems, including public safety networks that compete with or replace private cellular
UEs (User Equipment), NG-RAN and 5G Core network.
technologies such as TETRA, while providing much higher data rates,
What is network slicing? Unlike previous 3GPP mobile architectures
with full isolation w.r.t. other slices, and with maximum priority in the
where a single packet transport carrier was intended to carry all
event of a disaster (dynamic policy control). Another well-known target
types of services, in 5G the transport network is expected to provide
type of slice is the creation of multicast or broadcast networks.
optimized support for different types of services and users in terms of
Technical Specification TS22.261 [1] specifies a set of performance
throughput, latency, positioning, mobility, reliability, and availability.
This new evolution step requires very important changes both in the requirements for high data rate and traffic density scenarios, including
wireless and wired networks, including the scalable assignment of ITS (Intelligent Transport Systems) infrastructure backhaul. A sample of
network, computing and storage resources. the most demanding cases is summarized in Table 1. The table shows
Network slicing allows the operator to deploy a number of instances downlink (DL) and uplink (UL) rates, capacity in terms of 𝑏∕𝑠∕𝑘𝑚2 , la-
of virtual connectivity service tailored for specific purposes on top tency bounds, reliability as a percentage of correctly delivered packets
of the same physical infrastructure. The network slice features the can include several re-transmissions, the availability accounts for the
whole mobile connectivity function, including IMS (IP Multimedia percentage of time the service is up via its interface.
Subsystem); it can have specific functionality (e.g. priority, charging, Other requirements include timing, positioning, ranging (distance
policing, security, mobility), performance (e.g. latency, jitter, availabil- estimation between UEs), IoT (Internet of Things) traffic collection
ity, reliability and rate), or it can serve specific users (e.g. public safety capability, IMS multimedia telephony service, Machine-Learning-based-

∗ Corresponding author.
E-mail addresses: dlarra@it.uc3m.es (D. Larrabeiti), luismiguel.contrerasmurillo@telefonica.com (L.M. Contreras), gaoterop@it.uc3m.es (G. Otero),
jahgutie@it.uc3m.es (J.A. Hernández), juanpedro.fernandez-palaciosgimenez@telefonica.com (J.P. Fernandez-Palacios).
URL: https://www.it.uc3m.es (D. Larrabeiti).
1
By stage 2 is meant architectural description to fulfill requirements (stage 1), not implementation of the architecture (stage 3) (ITU-T Recommendation
I.130).

https://doi.org/10.1016/j.yofte.2022.103220
Received 6 May 2022; Received in revised form 24 December 2022; Accepted 31 December 2022
Available online 24 January 2023
1068-5200/© 2023 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
D. Larrabeiti et al. Optical Fiber Technology 76 (2023) 103220

Fig. 1. An example of management domains involved in a network slice.

prediction of QoS changes, tactile and multi-modal communication, Table 1


Sample service requirements from 3GPP TS22.261 [1].
security and charging.
Services such as URLLC (Ultra-Reliable Low-Latency Communica- Service Requirements

tions), connected industry 4.0 and V2X (Vehicle-to-everything), require Urban Downlink (DL) datarate: 50 Mbit/s; Uplink (UL) datarate:
25 Mbit/s; DL Capacity: 100 Gbit/s/km2 UL Capacity: 50
stringent QoS guarantees from the network. TS22.261 makes com-
Gbit/s/km2
pulsory to provide a mechanism to support real-time end-to-end QoS
Dense urban DL: 300 Mbit/s; UL: 50 Mbit/s; DL Capacity: 750
monitoring for a slice, including event notification to UEs or groups
Gbit/s/km2 ; UL Capacity: 125 Gbit/s/km2
of UEs. The goal is to enable some sort of error-reaction ability to
Indoor DL: 1 Gbit/s; UL: 500 Mbit/s; DL Capacity: 15 Tbit/s/km2 ;
interconnected automation devices by means of a continuous stream
UL Capacity: 2 Tbit/s/km2
of reports and asynchronous events upon crossing certain thresholds of
Rural DL: 50 Mbit/s; UL: 25 Mbit/s; DL Capacity: 1 Gbit/s/km2 ;
loss and/or latency. UL Capacity: 500 Mbit/s/km2
This means that, not only the operator needs to estimate, plan,
Broadcast DL: 200 Mbit/s/TV channel; UL: N/A or 500 kbit/user
deploy and monitor the performance of slices, but it must provide their
Tactile/Immersive Audio delay: 50 ms; Visual delay: 15 ms delay; Tactile
users with real-time information of the perceived end-to-end QoS, for
delay: 25–50 ms
which current multi-layer multi-technology networks are not prepared.
Wireless road-side Max. end-to-end latency: 30 ms; Survival time: 100 ms;
Furthermore, slicing needs to coexist with other high priority real-time infrastructure Service availability: 99.9999% ;Reliability; 99.999%
traffic like fronthaul traffic. backhaul
This article attempts to identify the challenges behind end-to-end Medical monitoring Max. end-to-end latency: < 100 ms; Service availability:
performance monitoring, sheds light over relevant recent steps toward > 99.9999%; MTBF (Mean Time Between Failures): >> 1
OAM convergence and provides an overview of latency planning tech- month
niques to deal with the question. To this end, the paper is structured Cloud/Edge rendering Max.end-to-end latency: 5 ms; Reliability: 99.99% (UL),
as follows. Section 2 describes the network technologies involved in a 99.9% (DL)
slice. Section 3 describes the OAM capabilities available in each of them Gaming/Interactive Max. end-to-end latency: 10 ms; Reliability: 99.99%
to contribute to a detailed latency analysis. Section 4 outlines the steps Virtual reality Max. end-to-end latency: 5–10 ms; Reliability: 99.99%
toward a unified mechanism that will allow a framework for end-to-end Split AI/ML image Max. UL end-to-end latency: 2 ms; Reliability: 99.9%;
performance monitoring. Section 5 proposes alternatives for latency recognition Service availability: > 99.999%
planning in the transport network and implementing the 5G slicing Split control for Max. DL end-to-end latency: 12 ms
architecture specification, Section 6 reviews the special case of latency robotics
management for fronthaul traffic (which has much more stringent AI/ML model Max. DL end-to-end latency: 1 s; Reliability: 99.9% for
requirements than regular slices) and Section 7 draws conclusions. A distribution for image transmission of model weight factors, 99.999% for
glossary of acronyms used throughout this article can be found in this recognition transmission of model topology
latter Section in Table 3.

2. Challenges behind end-to-end slice latency and jitter monitor- The amount of network elements involved in the service is very
ing for 5G large. Besides the classic layer 0 (photonic) to layer 3 (IP) architecture
(the figure shows ethernet switches, routers, ROADMs), as envisioned
When provisioning end-to-end slices for use cases of low mobility by 5G, it includes the forwarding the user traffic through data cen-
and guaranteed QoS, telecom operators need to estimate primary and ter servers (in the figure, 5G core and Multi-Access Edge Computing
backup paths with a maximum target latency. These paths can be (MEC)) and transit operators. Furthermore, carried traffic can contain a
packet-switched or, for especial applications, circuit switched. The mix of backhaul and fronthaul flows, the latter with stringent transport
service will be typically defined as a chain of PNF (Physical Network requirements. In order to deal with this complexity, telecom operators
Function) and VNF (Virtual Network Function) instantiations. One have split their networks into three management domains: RAN (Radio
major challenge in configuring, monitoring and troubleshooting slices Area Network), Transport and 5G core.2 The reason for this is the fact
comes from the intrinsic complexity of the end-to-end slice itself, made that each segment has its specific infrastructure, requirements and pro-
up of stitched service instantiations (most of them shared by sets of tocols; consequently, each one is usually managed by a different team
slices) across multiple network management domains each of them with different management tools. This makes end-to-end monitoring a
featuring a multiplicity of layers and technologies. This complexity is complex task. In this article, we set aside all QoS aspects related to the
illustrated in Fig. 1. The figure shows the transport network infrastruc- air interface and focus on the fixed, fiber-based side of the service.
ture involved in the delivery of a point-to-point network slice for an
industrial remote control use case. This use case has low latency, low
jitter and reliable connectivity as main requirements, and the novel 2
It should be noted that 3GPP names Access Network the infrastructure that
additional requirement by 5G of informing of the available QoS in real interconnects the User Equipment (UE) with the core, which usually spans
time. what operators call Access and the MAN.

2
D. Larrabeiti et al. Optical Fiber Technology 76 (2023) 103220

to the 5G core (5GC in the picture), and is the WAN technology


used beyond the 5GC to connect to the Internet or to other UEs in
a 5G network. In the Figure, the UE on the right is the controller
of the UE on the left and the Network Slice guarantees a given
QoS. Flex Ethernet (FlexE) has been identified as a useful tool
to implement slicing in the MAN and in the transport of D-
RAN traffic, to the point that its framing has been adopted in
Recommendation ITU-T G.8312. The challenge in this domain is
not at the optical layer given its circuit-switching nature, but at
the packet-switched layer above. The supervision of service level
objectives of network slices need to be performed both at the
IP layer and underlying technologies (MPLS-TP/Ethernet, FlexE).
On the other hand, when rate or latency bounds are not met, a
smart integration of service and photonic layers’ control planes is
Fig. 2. D-RAN elements, terminology and comparison with a conventional base station. indispensable to correct and adapt the number of optical channels
to the demand. In fact, the options are manyfold, from hard
isolation mechanisms (slices as wavelengths or as FlexE clients)
to soft isolation ones (packet-based bandwidth management in
The main characteristics of the domains and their challenges are
enhanced VPNs). Section 4 provides insight on how this integra-
the following:
tion has become doable in the light of SDN (Software Defined
• RAN domain. The RAN is made up of a combination of access Networking).
technologies ranging from PON (WDM-PON, NG-PON2) to point- • 5G Core Network domain
to-point fibers, Ethernet, FlexEthernet and MPLS-TP or IP/MPLS The design of the 5G Core (5GC) by 3GPP has pursued to leverage
or Segment Routing. Fig. 1 shows a scenario where the left UE the advantages of virtualization, with the possibility of instan-
(User Equipment) is connected to a 5G system implementing tiating the different constituent entities as virtual functions in
a distributed RAN (D-RAN) configuration. In other words, the distributed compute facilities across the network, in the form
base station is actually a RU (Remote Unit) implementing part of of virtual machines, containers, etc. The entity that provides
the radio signal processing chain, that forwards the digital radio connectivity at the data plane is known as User Plane Function
signal to the DU (Distributed Unit) in ethernet frames, where (UPF), which is in charge of terminating the Generic Transport
the MAC function is run. The network latency budget for this Protocol (GTP) tunnels established from the radio access node
transmission RU-DU (fronthaul) is 100 μs including propagation (Fig. 3). The access node can be the gNB (5G base stations)
delay [2]. Then packets are further forwarded to the CU (Central in the case of the traditional solution, or CU-UP (Central Unit-
Unit) with sub-ms budget (midhaul traffic), where the final data User Plane, the user plane part of the Packet Data Convergence
packets to be transported are generated (backhaul traffic). This Protocol) for Distributed RAN implementations (when radio func-
CU function may be co-located with the DU or be placed deeper tional split based architectures are deployed). The UPF can have
into the MAN (Metropolitan Area Network). In small MANs this an intermediate entity at the edge node (I-UPF) that allows to
DU/CU functionality may be fully centralized (C-RAN). enforce flow policies in the edge, besides serving as a switching
These three elements and the terminology are outlined in Fig. 2. point to other cores. Details on further 5GC entities can be found
The architecture can even support different placements of D-RAN in TS23.501 [3].
network functions for each service. Low latency services require The aforementioned 5GC entities can be deployed across multi-
the CU to be as close to the user as possible. Finally, this D-RAN ple distributed compute execution environments, interconnected
traffic may share the ethernet or IP infrastructure with backhaul through segments of the transport network for inter-cloud con-
traffic coming from gNBs, regular 5G base stations performing nectivity. The interconnection of those entities to the network,
RU/DU/CU functions in the base station. A microwave network including the UPF, implies the interaction of the logical connec-
segment may also be present as an alternative to optical fiber tions on the compute side with the physical connection of the
backhaul. The major challenge in this domain is fulfilling the compute infrastructure and the network elements in the transport
QoS of slices transported as backhaul while contending with network. In this context, implementing the UPF as microservices
high-priority fronthaul. IEEE802.1CM provides configurations for in a distributed compute environment in a latency-aware way as
Ethernet, where fronthaul has maximum priority. However not envisioned by 3GPP seems to be the most important challenge of
all existing RAN infrastructure is based on Ethernet switching, this domain to be met in the near future. The UPF implemented
and the impact of this priority scheme on lower priority traffic as a pure cloud native network function without dedicated hard-
can be important. Therefore, besides traffic engineering, proper ware can provide unprecedented levels of scalability, processing
monitoring mechanisms need to be in place along with scope for flexibility and re-programmability of user plane. However, the
action (e.g. planned alternative paths triggered by the monitoring development of such real-time systems capable of beating custom
applications) when high traffic conditions cause QoS violations. silicon is at an early stage.
There are OAM mechanisms to carry out such monitoring on a per
technology basis, as reviewed in the next section. This, together 3. The basic tools: relevant OAM protocols
with a progressive aperture of the control and OAM interfaces of
devices, is expected to enable a proper end-to-end latency control, Continuous QoS monitoring of slices is necessary because any cost-
as implicitly required by the 5G slice concept. effective network design needs to rely on statistical multiplexing, which
• Transport Network domain. The transport network domain is a usually leads to statistical guarantees. Moreover, the operator needs
multi-layer IP/WDM infrastructure in large settings or IP/Ethernet accurate real-time per-segment knowledge of the latency sources. Al-
in smaller ones. The IP/WDM transport technology may be OTN though, as the previous section showed, this is quite challenging,
(Optical Transport Network ITU-T G.709), usually transporting the fact that most technologies have their own latency measurement
Ethernet line signals between router interfaces in an ODU carrier. mechanisms and OAM protocols makes it viable. A quick review of this
The transport network spans the MAN, from the 5G edge nodes capability in currently involved technologies follows:

3
D. Larrabeiti et al. Optical Fiber Technology 76 (2023) 103220

Fig. 3. The user plane function in the 5G architecture.

• 5G Access Network. 3GPP TS29.244 [4] at stage 3 describes the • Carrier Ethernet. A complete framework for Connectivity Fault
Implementation of the interface between the Control Plane and Management (CFM) protocols exist: IEEE 802.1ag and ITU-T
the User Plane Nodes. It specifies the support by 5G systems of Y.1731. Regular in-band OAM frames allow the measurements,
QoS monitoring and defines the signaling to setup a reporting which include frame delay and frame delay variation measure-
period or programming reports when the delay exceeds a thresh- ment for multi-hop flows between Maintenance Points, based on
old. Measurements are taken from the UE to either the UPF(PSA) configured Maintenance Associations. IEEE 802.3-2005 (previ-
or the I-UPF (Fig. 3). The former is the entity connected to the ously 802.3ah, Clause 57) includes Ethernet link OAM for a single
transport data network. The latter is an optional entity located at link, which supports delay measurement only in loopback mode.
the MEC (Multi-Access Edge Computing) edge data center. The • FlexEthernet. FlexE was proposed by the Optical Internetwork-
measurement can be uplink, downlink or round trip. This means ing Forum (OIF), as a mechanism to decouple MAC and PHY
that the mechanism to convey a session’s QoS in real time (either layers of Ethernet clients; it features bonding, sub-rating, and
for the UE or per flow) exists, for the segments UE-5G Core. The channelization of 1 to 𝑚 100GBase-R PHYs (200G and 400G
breakdown of latencies in those segments and in the transport in the future) and can be used in Router to Transport connec-
networks should be obtained and monitored by other means, as tion scenarios where the mapping/de-mapping FlexE Shim layer
described in the next section. At provisioning time the operator allows to flexibly partition and assign bandwidth groups of 5
may run standard testing procedures such as the Two-Way Active Gb/s slots to individual flows [8]. As mentioned above, FlexE
Measurement Protocol (TWAMP) (RFC5357) to check that the has been identified as an alternative to implement slicing in the
slice fulfills the target performance end to end at the IP layer. MAN and in the transport of D-RAN traffic. As an example of this
• OTN and DWDM channels. OTN ITU-T Rec. G.709/Y.1331 sup- application, in [9], a number of 5 Gb/s slots was calculated for the
ports round-trip time measurements between Path Connection transport of different 5G New Radio configurations, showing that
Monitoring End Points by means of Delay Measurement (DM) bits latency can be bounded if the FlexE groups are properly dimen-
in the ODU overhead. The corresponding remote End Point is sioned. ITU-T G.8312 Recommendation adopting FlexE describes
programmed to copy back the DM bit it receives in the returning its own performance monitoring mechanism that includes delay
ODU frames, in such a way that whenever the initiator End Point measurement.
changes the DM bit, the round-trip time can be estimated by • MPLS, MPLS-TP. For MPLS, ITU-T Recommendations Y.1710 and
the time the change is detected back at the initiator End Point. Y.1711 deal with fault management, support OAM frames, but
The resolution achieved is the duration of two OTU frames, for do not include delay measurements. However, ITU-T Recommen-
instance, at 100 Gb/s the error margin is 2.6 μs. This measurement dation ITU-T G.8113.1/Y.1372.1 (2012) does support proactive
can be performed on a per tandem connection basis, up to six and on-demand delay measurements for MPLS-TP. The same man-
nested tandem connections. Similar approaches can be followed agement entities of Ethernet layer networks (adopted in ITU-T
in non-OTN DWDM systems at the link layer (e.g. Ethernet link G.8010) are used in this recommendation, what makes conver-
OAM). gence easier. One-way delay measurements are supported if ac-
• Passive Optical Networks. In general, TDMA (Time Division curate time synchronization is available at both Maintenance End
Multiple Access) PONs have issues in guaranteeing short uplink Points.
access delay, unless specific DBA (Dynamic Bandwidth Alloca-
tion) protocols can guarantee TDM-based reservations with fixed All these technology-specific OAM mechanisms can be leveraged
transmission slots within the 125 μs-frame so that it is possible to create monitoring mechanisms able to breakdown the end-to-end
to reduce the uplink access time as much as desired [5,6]. In latency into layers and domains. The goal is being able to automatically
general, the bandwidth requirements of fronthaul traffic also locate the network elements responsible for a latency noncompliance.
demand upstream capacities of 10 Gb/s and even more to comply
with reduced upstream delay percentiles [7]. This is definitely 4. Multi-layer multi-domain OAM convergence efforts
a must to transport fronthaul traffic, given the 100 μs budget,
or any other sub-ms service. WDM-PON standards overcome the The split of network management into domains described in Sec-
jitter issues of TDMA PON. WDM-PON can properly work with tion 2 and the contradictory need to have real end-to-end control of
IEEE 802.1ag OAM as they provide point-to-point links usually QoS parameters by 5G make necessary to create a converged man-
exploited by ethernet transceivers. TDM-PON have embedded agement layer on top of a service bus where all the management
OAM mechanisms dealing with connectivity fault management; domains are present. This idea as depicted in Fig. 4. The Figure de-
however, the main TDM-PON standards lack proper per-class QoS scribes how the overall network and service management system em-
monitoring support and latency measurement from an upper layer beds all the management domains to provide a view of the end-to-
is required, what makes WDM-PON a better option in this regard. end service with detailed information of the latency components. The

4
D. Larrabeiti et al. Optical Fiber Technology 76 (2023) 103220

Fig. 4. Convergence of management domains and layers toward real-time latency monitoring and dissection.

picture shows the RPCs issued by the respective domains to man- allowing to present timely information about the compliance of the
agement end points at the three layers: layer 2 (ethernet), layer 3 committed SLOs to the customer requesting the network slice.
(IP/MPLS, IP/MPLS-TP, IP/Segment Routing) and the circuit-switched It is the mission of the NSC to determine the technology for the slice
optical layer (TDM/WDM channels, OTN, FlexE/WDM, FlexE/OTN). realization, based on the SLOs and the connectivity matrix defined by
The system takes advantage of the OAM capabilities existing at each the customer. With that purpose, the NSC, once the request has been
layer to measure delay, but they are invoked through standard proto- processed, will trigger the provisioning process directly interacting
cols rather than proprietary ones. with the aforementioned per-technology domain SDN controllers. In
Indeed, the progressive endorsement by operators and manufactur- the case of the optical domain, the interface of reference for such
ers of open multilayer OAM solutions [10] is enabling integrated and interaction is the Transport API (TAPI). TAPI [13] is a standard API
fully detailed view of latency across layer segments with SDN philos- developed by the Open Networking Foundation (ONF) that allows a
ophy. Connectivity and latency supervision tools of most technologies TAPI client (e.g., an orchestrator as described before) to control a
are becoming available to controllers through NETCONF or RESTCONF, domain of transport network equipment controlled by multiple TAPI
as equipment manufacturers keep opening and standardizing their servers (e.g. a specific per-technology domain controller). TAPI allows
interfaces. For example, In 2017, IEEE, MEF, and ITU-T SG15 started the control of network resources at different levels of abstraction, re-
a liaison toward a IEEE 802.1 CFM YANG data model, whose result is sembling TMN’s layered approach (Telecommunications Management
Network). Given the importance of considering Operation and Main-
IEEE 802.1Qcp-2018, available at GitHub.
tenance (OAM) mechanisms to ensure the proper behavior of the
When considering the request and instantiation of a network slice
allocated resources and connectivity resources, TAPI introduced the na-
at the transport side, in line with the architecture depicted in Fig. 4, it
tive support of OAM services. For instance, the TAPI 2.0 model has been
can be expected the presence of a hierarchical structure where the top
extended to support Maintenance Entities, Maintenance Entity Groups,
SDN controller is able to orchestrate different per-technology domain
Maintenance End Points and Maintenance Domain Intermediate Points,
SDN controllers below. This is for instance the proposed architecture
in line with the ITU-T Y.1731/IEEE 802.1ag specification. This allows
in relevant projects such as the Telecom Infra Project MUST [10]. The
the TAPI client to determine where monitoring points exist along a
hierarchical SDN controller is expected to support the functionality connection and to launch measurement cycles between them.
of Network Slice Controller (NSC), as defined in [11]. This NSC is TAPI follows a multi-layer and multi-technology approach (Fig. 4),
expected to offer at its Northbound Interface (NBI) a data model able including optical but also Ethernet technologies in scope. However, it is
to support network slices requests from different customers, as could being mostly positioned as an interface for optical networks, as in [14].
be the case of an upper management system as the one from 3GPP In this case, when interacting with the NSC above, the NSC will instruct
or others. The data model conceived for the NSC [12] intends to be the optical controller to set up a number of optical paths compliant with
a technology-agnostic model in which the network slice is requested the latency expectations informed by the customer through the network
expressing the needs in terms of connectivity and the associated Service slice NBI, as well as the necessary OAM measurements to ensure that
Level Objectives (SLOs), but there are no indications on how the the service performance observed matches the customer expectations.
network slice should be realized in terms of the technology to be used Another relevant contribution which is agnostic of the control
for that purpose. The SLOs will allow to indicate constraints for the framework is IETF LIME (Layer Independent OAM Management in
connectivity constructs between endpoints, including latency, jitter, the Multi-Layer Environment). This working group produced three
bandwidth, etc. This will be the base for determining later on the RFCs (IETF Request For Comments), currently Proposed Standard:
planning of the latency component for the transport part of the end-to- two YANG data models for OAM protocols connection oriented and
end network slice. It should be noted that such data model also includes connectionless (RFC8531, RFC8532), and a retrieval method YANG
monitoring parameters for measurable SLOs such as the ones described, data model for connectionless OAM (RFC8533). RFC8533 provides

5
D. Larrabeiti et al. Optical Fiber Technology 76 (2023) 103220

technology-independent RPC (Remote Procedure Call) operations for


OAM protocols that use connectionless communication, extensible with
technology-specific details. Fig. 4 tries to illustrate this concept. Each
Management Domain issues RPCs (depicted by vertical arrows) to
different Network Configuration Protocol (NETCONF) servers in or-
der to execute secure OAM commands which are multi-layer and
multi-technology. The horizontal arrows describe the breakdown of
latency components at each level. The ‘‘connectionless-oam-methods’’
of module RFC8533 defines RPC ‘continuity-check’, similar to IP ping
(RFC792, RFC4443) and MPLS LSP ping (RFC8029), and
‘path-discovery’ (equivalent to IP traceroute) operations.
Finally, it is worth mentioning other relevant research work pro-
viding novel ideas to support real-time slice latency monitoring not Fig. 5. Latency components for a distributed service of 𝐾 latency-bounded stages.
relying on existing OAM protocols. The authors of [15] propose to
equip network slices with an end-to-end latency sensor as an additional
function in the service chain. These sensors further allow to optimize
100 Gb/s) and propagation (this is 5 μs per Km of fiber) delay compo-
latency through path reconfiguration of optically interconnected data
nents. In total, let us consider that each IP hop contributes with a delay
centers.
which is a random variable X𝑖 whose probability density function is
unknown but bounded between 𝑎 and 𝑏 with a mean 𝐸(X𝑖 ), 𝑖 = 1, … , 𝐾.
5. Planning and configuring bounded latency for a slice In this scenario, we can apply the Hoeffding’s inequality which states
the following: ‘‘Suppose X1 , … , X𝐾 are independent random variables
Estimating and planning the latency of slices can be a complex task ∑
taking values in [𝑎, 𝑏], and let denote their sum X = 𝐾 𝑖=1 X𝑖 with mean
due to the length of the transmission paths crossing lots of switching value 𝜇 = 𝐸(X)’’. Then for any value of 𝛿 > 0:
and data processing devices that make up the so-called VNF service
2𝛿 2 𝜇 2
chain. The authors of [16] propose and demonstrate the use of a −
P (X > (1 + 𝛿)𝜇) < 𝑒 𝐾(𝑏−𝑎)2 .
Latency-Aware Service Chain Computation Element (LA-SCCE) as an
evolution of the classical Path Computation Element (PCE) tuned for For instance, if all aforementioned latency components of each IP
service chain allocations, where the path is constrained to traverse a se- hop are modeled as a random variable bounded within [5, 50] μs with
quence of VNFs with latency requirements. Indeed, once the maximum mean 20 μs in a 100 Gb/s MAN path, this formula lets us check that
processing times are known to the network designer it is possible to at least 99.986% of the packets will traverse the path under a target
allocate resources smartly, keeping latency under control. However the latency budget of 0.5 ms if the path has 𝐾 = 10 hops. However, this
time spent by packets traversing services running on Virtual Machines guarantee is reduced to 92.8% for a 15-hop path.
(VMs) of data centers can vary a lot (due to VM mobility, availability Beyond giving a practical estimation of percentiles of conventional
of parallel resources, NFV, etc.) and sufficient delay margins need to packet forwarding latency, Hoeffding’s inequality is particularly useful
be considered in the design at the transport level. in the context of network slices for chained NFVs or any type of
The following simplified analysis provides a set of numerical exam- distributed packet processing service whose latency components are
ples on delay bounds that can be used as reference rules for designing bounded. For example, let us consider a network slice for an AI/ML
network slices with latency guarantees. This should be considered as image recognition service pipelined throughout 𝐾 processing nodes
a set of simplified guidelines, rather than a precise design procedure. with a total end-to-end delay requirement of 2 ms (see Table 1), and,
As an example, consider a network slice which requires that the 99th for simplicity, let us assume that the processing can be performed on a
percentile delay is strictly below 250 μs, i.e., 𝑑.99 = 250 μs. In other per-packet basis (e.g. fingerprint matching). This scenario is illustrated
words, this means that 99% of the packets must experience a latency in Fig. 5, where each node is known to cause a variable delay between
value of 250 μs or less. Similarly, 𝑑.95 or 𝑑.90 may be used as a design [0.1, 0.5] ms with an average value of 0.2 ms.
criteria. Additionally, let us consider a path traversing 𝐾 links from Let us assume that the number of processing hops 𝐾 = 5. Then
2 ms
source to destination, where each link carries traffic of at most 50% of 2 ms = (1 + 𝛿)𝜇, leading to a value of 𝛿 = 5×0.2 ms
− 1 = 1. Hence,
its capacity (i.e. 50% link-load) to provide spare capacity for backing the Hoeffding’s inequality states that:
up other connections. In particular, we consider the scenario of Fig. 4, 2 2
− 2⋅1 ⋅12
between the CU/DU and the 5G Core. P(X > 2 ms) = P(𝑋 > (1 + 1) ⋅ 5 ⋅ 0.2 ms) < 𝑒 5⋅0.4 = 0.082
Following the Kingman’s approximation for generic G/G/1 queue- which means that, in such situation, less than 8.2% of the packets
ing models, the average waiting time in queue 𝐸(𝑊𝑞 ) of a packet experience delays higher than 2 ms. In other words, 2 ms would
selected at random in a link with load 𝜌 follows [17] and can be represent the 91.8th delay percentile (i.e. 𝑑.918 = 2 ms). Finally, Table 2
expressed as: shows the application of Hoeffding’s inequality for different use-case
𝜌 𝐶 2 [T] + 𝐶 2 [S] scenarios. It is worth noting that when the number of IP hops is reduced
𝐸(W𝑞 ) ≤ 𝐸[S] , to 𝐾 = 3, almost all packets are guaranteed delays below 2 ms. If this
1−𝜌 2
is not possible, the latency in the processors needs to be improved.
where 𝐸[S] is the average service time of a packet, and 𝐶 2 [T], 𝐶 2 [S] are
Alternatively, relaxing the delay requirements of the service to 5 ms
the coefficient of variation of both packet arrivals and service times.
ensures that almost 100% of the packets are within the delay bound.
For instance, 𝐸[S] = 0.12 μs for 1500-byte packets transmitted over
100 Gb/s links; also 𝐶 2 [T] and 𝐶 2 [S] are equal to 1 for Poisson traffic
6. The case of fronthaul traffic
with exponentially-distributed service times. In such a case:
0.5 1 + 1 A special type of traffic deserving careful latency engineering is
𝐸(W𝑞 ) ≤ 𝐸[S] = 𝐸[S] = 0.12 μs
1 − 0.5 2 fronthaul traffic. As discussed in Section 2, the radio access network
The total delay per hop must take into account, not only queueing is evolving toward disaggregated and distributed approaches. D-RAN
delay, but also processing (between 0.1 μs and 2 μs for high-speed technologies come at the expense of high bandwidth utilization and
switches), transmission (another 0.12 μs for a 1500 byte packet at stringent delay and jitter requirements [18]. Standardization bodies

6
D. Larrabeiti et al. Optical Fiber Technology 76 (2023) 103220

Table 2
Percentage of packets whose delay is below a target budget
for 𝐾 hops. Each hop’s delay is considered a random variable
bounded within [0.1, 0.5] ms with mean 0.2 ms.
Delay budget → 2 ms 5 ms
K = 3 hops 99.97% ∼ 100%
K = 5 hops 91.8% ∼ 100%
K = 7 hops 47.4% ∼ 100%
K = 9 hops 5.4% 99.99%

have defined different methods for implementing the D-RAN concept


in packet-based transport networks. It is worth highlighting the work
of the IEEE Time-Sensitive Networking (TSN) working group to support
the transport of time-sensitive fronthaul data over Ethernet networks.
The use of Ethernet networks leverages the high penetration of low-cost
Ethernet hardware along with the statistical multiplexing gains offered
by packet-switched networks. In the IEEE 802.1CM standard [2,19],
the TSN group gives important recommendations for the configuration
of the transport ethernet network to enable a successful transmission Fig. 6. Estimated vs simulated queueing delay in a fronthaul aggregator.
of fronthaul data. Mainly, it defines two different classes of fronthaul
interfaces: Class 1 for interfaces in which the signal processing chain
is decomposed according to the CPRI standard [20], and Class 2 for
for the end-to-end latency planning. Namely, in [23] we estimate the
those decompositions of E-UTRA base station in the intra-PHY splits,
average aggregated queueing delay assuming a G/G/1 queueing model.
i.e., Splits 𝐈U and 𝐈𝐈D . Additionally, IEEE 802.1CM suggests several tim-
In [17] we further extend the work to give a closed expression of the
ing distributions to fulfill the synchronization requirements of different
𝑝th percentile queueing delay for the aggregation of eCPRI fronthaul
features (handovers, MIMO, CoMP, etc.). These are (a) High Priority
Fronthaul (HPF), including Class 1 IQ data and Class 2 User Plane data, traffic,
{ }
with 100 μs maximum end-to-end one-way latency, (b) Medium Priority 1 𝐶 2 [T] + 𝐶 2 [S] 𝜌
W(𝑝)
𝑞 = max 0, 𝐸[S] 1 − 𝜌 ln( ) ,
Fronthaul (MPF), including Class 2 User Plane slow data, and Control 2 1−𝑝
and management fast data, with 1 ms one-way latency budget, and where S and T are random variables modeling the service time and
(c) Low Priority Fronthaul (LPF), for Class 1 and Class 2 control and
interarrival times of packets at the queue, and 𝜌 is the queue occupancy.
management data.
𝐶 2 [X] stands for the squared coefficient of variation of a random
Fronthaul transport may also be seen as a particular form of net-
variable, e.g., X and is defined as 𝐶 2 [X] = Var[X]
E[X]2
. From there, we derive
work slice with the most stringent requirements of all the transported
a set of rules for Ethernet-based fronthaul network dimensioning, using
traffic. Although fronthaul traffic needs to co-exist with network slicing
high delay percentiles as the key design metric, instead of conventional
traffic in mixed backhaul-fronthaul (xHaul) environments, it should be
average delays. As a numerical example, consider the output port of
remarked that fronthaul connectivity is not a 5G slice per se, as no
a concentrator aggregating fronthaul traffic with 70% load (𝜌 = 0.7).
functionality to a subset of UEs is realized. In fact a fronthaul flow
Assuming T =≃ 8.8 and S = 0 as typical values characterizing the
normally carries the traffic of multiple slices.
fronthaul traffic (see [17]), we get that the 90th percentile of the
If the slice is transported within the fronthaul flows, the stringent
queueing delay is W(90th)
𝑞 = 34.25 μs, which is roughly 30 times
requirements of the fronthaul traffic are applied to those services on
higher than the average value and significant compared with the total
top of the slice while they are within the fronthaul network. The
aforementioned budget of 100 μs.
envisioned end-to-end latency for future 5G and beyond-5G networks
depends on the application. Most will require latencies below a few In Fig. 6, we apply this methodology to the aggregation of fron-
milliseconds [21,22], for instance, in tactile internet, factory automa- thaul flows carrying 20 MHz LTE channels. We compare the estimated
tion (≤ 1 ms), intelligent transportation systems (≤ 4 ms), etc. This queueing delay and the simulated values as we increase the aggregated
translates into even more strict requirements at the lower layers of traffic. Finally, it is worth mentioning a useful fronthaul network
the signal processing chain if these services are transported in the dimensioning tool [24] for those cases where the distance between the
fronthaul traffic. Regarding the fronthaul itself, the end-to-end Ethernet antennas and the processing units needs to be maximized. Based on
network latency target is 100 μs for CPRI as defined in 802.1CM [19]. the use of very high packet delay percentiles, this method represents
This is a useful design parameter which shows the importance of the an alternative to the use of maximum theoretical delay as the main
characterization of the queueing delay throughout the network. By dimensioning metric. By interpreting the gap between the extreme per-
keeping the queueing delay low, the budget for propagation and fabric centiles and the maximum worst-case delay as an extra delay budget,
switching delays can be expanded, increasing the reach of the network. the fronthaul links can be further stretched while complying with the
Some additional delay budget can be gained by embracing higher delay and frame loss ratio defined in IEEE 802.1CM. Experiments show
functional splitting of the signal processing chain, e.g., Intra-PHY or that the fronthaul links’ lengths can be increased by 60% and 10% for
MAC-PHY splits, since these have more relaxed synchronization, delay, 50 MHz and 100 MHz 5G New Radio channels, respectively.
and jitter requirements. It should be noted that the analysis addresses On the other hand, if the slices are transported as backhaul traffic,
only the packet transport latency, i.e. the time since the first bit of the they might share the same transport network and compete with fron-
packet is sent over the network interface until the last bit of the packet thaul flows for the network resource. In this case, special attention must
is delivered at the remote interface. Additional latency components be put to the network planning so as to guarantee that the requirements
exist at the endpoints, namely jitter compensation delay (matching the of both the fronthaul and backhaul traffics are met. In general, the
maximum queueing delay) and OFDM’s symbol reassembly time (since worst-case capacity available for backhaul should be decreased by the
it travels in a burst of packets). sum of all fronthaul flows at maximum load. Then the latency for a slice
In some of our previous research, we studied the behavior of the packet should be additionally increased by the maximum aggregate
aggregated fronthaul traffic and developed some tools that are useful burst size of fronthaul flows converging on the output line. In previous

7
D. Larrabeiti et al. Optical Fiber Technology 76 (2023) 103220

Table 3 static latency planning both in packet networks carrying slices and
Glossary.
fronthaul traffic. However, real-time monitoring enables smart dynamic
3GPP 3rd Generation Partnership Project https://www.3gpp.org/ reconfiguration of network resources. The trend to open OAM interfaces
5GC 5G Core
API Application Programming Interface
is a major advance toward this goal. By adopting a unified OAM
CU Central Unit approach, recent works on data-driven dynamic resource scheduling
CFM Connectivity Fault Management mechanisms, such as [26,27], gain practical relevance. These works aim
DWDM Dense Wavelength Division Multiplexing to optimize slice configurations upon the analysis of network usage and
IMS IP Multimedia Subsystem
performance data with techniques such as deep reinforcement learning.
ITS Intelligent Transportation Systems
IoT Internet of Things
IP Internet Protocol CRediT authorship contribution statement
DU Distributed Unit
FlexE Flexible Ethernet https://www.oiforum.com/
LSP Labeled Switched Path David Larrabeiti: Conceptualization, Writing – original draft, Writ-
MAN Metropolitan Area Network ing – review & editing. Luis M. Contreras: Conceptualization, Writing
MEC Multi-Access Edge Computing – original draft, Writing – review & editing. Gabriel Otero: Writing –
MEF https://www.mef.net/ review & editing, Validation. José Alberto Hernández: Writing – orig-
MPLS Multiprotocol Labeled Switching
MPLS-TP MPLS Transport Profile https://www.itu.int/rec/T-REC-G.8110.1/en
inal draft, Writing – review & editing. Juan P. Fernandez-Palacios:
NBI Northbound Interface Conceptualization, Investigation.
NETCONF Network Configuration Protocol RFC4741
NG-RAN Next-Generation Radio Area Network
Declaration of competing interest
NG-PON2 Next-Generation PON version 2
https://www.itu.int/rec/T-REC-G.989
NSC Network Slice Controller The authors declare that they have no known competing finan-
OAM Operations, Administration and Management cial interests or personal relationships that could have appeared to
OTN Optical Transport Network ITU-T Rec. G.709
influence the work reported in this paper.
PCE Path Computation Element
PNF Physical Network Function
PON Passive Optical Network Data availability
QoS Quality of Service
RAN Radio Area Network
RPC Remote Procedure Call
No data was used for the research described in the article.
RU Remote Unit
ROADM Reconfigurable Optical Add-Drop Multiplexer Acknowledgments
SDN Software Defined Networking
SLO Service Level Objective
TAPI Transport API This work has been partially funded by the EU H2020 project
TETRA TErrestrial Trunked RAdio https://www.etsi.org/technologies/tetra Int5Gent (grant no. 957403), as well as the NextGenerationEU project
TMN Telecommunications Management Network ITU-T Recommendation 6G-Xtreme (grant no. AEI/10.13039/501100011033), TAPIR-CM (grant
series X.700
no. P2018 /TCS-4496) and ACHILLES (grant no. PID2019-104207RB-
TWAMP Two-Way Active Measurement Protocol
UE User Equipment (5G terminals)
I00). Funding for APC: Universidad Carlos III de Madrid (Read &
URLLC Ultra-Reliable Low-Latency Communications Publish Agreement CRUE-CSIC 2022)
UPF User Plane Function
V2X Vehicle-to-everything
References
VM Virtual Machine
VNF Virtual Network Function
YANG Data Modeling Language for NETCONF RFC6020 [1] 3GPP, Service requirements for the 5G system; stage 1, Technical Spec-
ification (TS), (21.261) 3rd Generation Partnership Project (3GPP), 2022,
URL https://portal.3gpp.org/desktopmodules/Specifications/SpecificationDetails.
aspx?specificationId=3107. version 18.6.0.
[2] IEEE802.1CMde, IEEE 802.1cmde-2020 - time-sensitive networking for fronthaul
studies [25], we proposed a distributed mechanism to orchestrate the amendment: Enhancements to fronthaul profiles to support new fronthaul inter-
end-to-end network resources to comply with the various requirements face, synchronization, and syntonization standards, 2020, URL https://standards.
of different services in a heterogeneous C-RAN network, including ieee.org/standard/802_1CMde-2020.html.
cellular and machine-to-machine communications, MEC, and fronthaul [3] 3GPP, System architecture for the 5G System (5GS); Stage 2 (Release 16),
Technical Specification (TS), (23.501) 3rd Generation Partnership Project (3GPP),
transport services in a converged architecture.
2022, Version 16.12.0.
[4] 3GPP, Interface between the Control Plane and the User Plane Nodes; Stage 3
7. Conclusions (Release 17), Technical Specification (TS), (29.244) 3rd Generation Partnership
Project (3GPP), 2022, Version 17.4.0.
Currently, most 3GPP specifications relevant to 5G slicing are at [5] R. Bonk, T. Pfeiffer, New use cases for PONs beyond residential services, in:
2022 Optical Fiber Communications Conference and Exhibition, OFC, 2022, pp.
stage 2. One main exception is 3GPP TS29.244 [4]. At stage 3, it
1–3.
describes the Implementation of the interface between the Control and [6] D. Eugui, J.A. Hernández, Analysis of a hybrid fixed-elastic DBA with guaranteed
User Plane Nodes, mandating QoS monitoring and detailing the signal- fronthaul delay in XG(s)-PONs, Comput. Netw. 164 (2019) 106907, http://
ing to be used to activate QoS reporting for a slice. However, there dx.doi.org/10.1016/j.comnet.2019.106907, URL https://www.sciencedirect.com/
are a lot of open research issues, mainly those regarding scalability science/article/pii/S1389128619302117.
[7] J.A. Hernandez, A. Ebrahimzadeh, M. Maier, D. Larrabeiti, Learning EPON delay
of resource reservations in the transport network and isolation among
models from data: a machine learning approach, J. Opt. Commun. Netw. 13 (12)
slices in statistically multiplexed resources, before the implementation (2021) 322–330, http://dx.doi.org/10.1364/JOCN.437414.
of 5G slicing becomes a reality. [8] A. Eira, A. Pereira, J. Pires, J. Pedro, On the efficiency of flexible ethernet client
The need for sharing resources among slices to achieve cost-effect- architectures in optical transport networks, J. Opt. Commun. Netw. 10 (1) (2018)
A133–A143.
iveness enforces the use of continuous performance monitoring on a
[9] J.A. Hernández, G. Otero, D. Larrabeiti, O.G. de Dios, Dimensioning flex ethernet
per-slice basis. This allows to perform real-time verification of service groups for the transport of 5G NR fronthaul traffic in C-RAN scenarios, in: 2021
levels and react quickly to non-compliances by network orchestrators. International Conference on Optical Network Design and Modeling, ONDM, 2021,
This paper provides some practical formulas to guide rule-of-thumb pp. 1–3, http://dx.doi.org/10.23919/ONDM51796.2021.9492417.

8
D. Larrabeiti et al. Optical Fiber Technology 76 (2023) 103220

[10] Telecom Infra Project, Open transport SDN architecture whitepaper, 2018, [19] IEEE802.1CM, IEEE standard for local and metropolitan area networks: Ieee time-
URL https://cdn.brandfolder.io/D8DI15S7/at/jh6nnbb6bjvn7w7t5jbgm5n/ sensitive networking for fronthaul 802.1cm, 2018, URL https://standards.ieee.
OpenTransportArchitecture-Whitepaper_TIP_Final.pdf. org/standard/802_1CM-2018.html.
[11] A. Farrel, J. Drake, R. Rokui, S. Homma, K. Makhijani, L. Contreras, J. Tantsura, [20] Industry cooperation: Ericsson, N. Huawei, Nokia, Common public radio interface
Telecom infra project. Framework for IETF network slices, 2022. (cpri), interface specification v7.0, 2018, URL http://www.cpri.info/spec.html.
[12] B. Wu, D. Dhody, R. Rokui, T. Saad, L. Han, Telecom infra project. IETF network [21] O.N.C. Yilmaz, Ultra-reliable and low-latency 5G communication, 2016.
slice service YANG model, 2022. [22] J.G. Andrews, S. Buzzi, W. Choi, S.V. Hanly, A. Lozano, A.C.K. Soong, J.C.
[13] Open Networking Foundation (ONF), The ONF Transport API (TAPI) project SDK. Zhang, What will 5G be? IEEE J. Sel. Areas Commun. 32 (6) (2014) 1065–1082,
URL https://github.com/OpenNetworkingFoundation/TAPI/releases. http://dx.doi.org/10.1109/JSAC.2014.2328098.
[14] Arturo Mayoral Lopez de Lerma, et al., Telecom Infra Project. MUST Optical [23] G.O. Pérez, J.A. Hernández, D.L. López, Delay analysis of fronthaul traffic in 5G
SDN Controller NBI Technical requirements, URL https://cdn.brandfolder. transport networks, in: 2017 IEEE 17th International Conference on Ubiquitous
io/D8DI15S7/as/557f4z3n738v4cww28qjxh6/MUST_Optical_Controller_NBI_ Wireless Broadband, ICUWB, 2017, pp. 1–5, http://dx.doi.org/10.1109/ICUWB.
Requirements_Document_v10_FINAL_VERSION_WEBSITE.pdf. 2017.8250956.
[15] R. Montero, F. Agraz, A. Pagès, S. Spadaro, End-to-end network slicing in support [24] G. Otero Pérez, D. Larrabeiti López, J.A. Hernández, 5G new radio fronthaul
of latency-sensitive 5g services, in: A. Tzanakaki, M. Varvarigos, R. Muñoz, network design for eCPRI-IEEE 802.1CM and extreme latency percentiles,
R. Nejabati, N. Yoshikane, M. Anastasopoulos, J. Marquez-Barja (Eds.), Optical IEEE Access 7 (2019) 82218–82230, http://dx.doi.org/10.1109/ACCESS.2019.
Network Design and Modeling, Springer International Publishing, Cham, 2020, 2923020.
pp. 51–61. [25] G.O. Pérez, A. Ebrahimzadeh, M. Maier, J.A. Hernández, D.L. López, M.F. Veiga,
[16] F. Moreno-Muro, C. San-Nicolás-Martínez, M. Garrich, P. Pavon-Marino, O.G. Decentralized coordination of converged tactile internet and MEC services in H-
De Dios, R.L. Da Silva, Latency-aware optimization of service chain allocation CRAN fiber wireless networks, J. Lightwave Technol. 38 (18) (2020) 4935–4947,
with joint vnf instantiation and SDN metro network control, in: 2018 European http://dx.doi.org/10.1109/JLT.2020.2998001.
Conference on Optical Communication, ECOC, 2018, pp. 1–3, http://dx.doi.org/ [26] H. Wang, Y. Wu, G. Min, J. Xu, P. Tang, Data-driven dynamic resource scheduling
10.1109/ECOC.2018.8535492. for network slicing: A Deep reinforcement learning approach, Inform. Sci.
[17] G.O. Pérez, J.A. Hernández, D. Larrabeiti, Fronthaul network modeling and 498 (2019) 106–116, http://dx.doi.org/10.1016/j.ins.2019.05.012, URL https:
dimensioning meeting ultra-low latency requirements for 5G, J. Opt. Commun. //www.sciencedirect.com/science/article/pii/S0020025519303986.
Netw. 10 (6) (2018) 573–581, http://dx.doi.org/10.1364/JOCN.10.000573. [27] B. Han, H.D. Schotten, Machine learning for network slicing resource man-
[18] O-RAN, Xhaul Transport Requirements XTRP-REQ-v01.00 , Technical Report, agement: A comprehensive survey, 2020, CoRR abs/2001.07974. URL https:
(29.244) O-RAN, 2021, Version 17.4.0. //arxiv.org/abs/2001.07974, arXiv:2001.07974.

You might also like