You are on page 1of 8

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/353490533

Dynamic Slice Scaling Mechanisms for 5G Multi-domain Environments

Conference Paper · June 2021


DOI: 10.1109/NetSoft51509.2021.9492716

CITATIONS READS

2 209

8 authors, including:

David Breitgand Alexios Lekidis


IBM Research -- Haifa Intracom Telecom S.A.
68 PUBLICATIONS   1,909 CITATIONS    38 PUBLICATIONS   262 CITATIONS   

SEE PROFILE SEE PROFILE

Rasoul Behravesh Vasileios Theodorou


Fondazione Bruno Kessler Intracom Telecom S.A.
16 PUBLICATIONS   72 CITATIONS    33 PUBLICATIONS   245 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Riabiligame View project

Small cEllS coordinAtion for Multi-tenancy and Edge services (SESAME) View project

All content following this page was uploaded by Alexios Lekidis on 23 August 2021.

The user has requested enhancement of the downloaded file.


Dynamic Slice Scaling Mechanisms for 5G
Multi-domain Environments
David Breitgand ⇤ , Alexios Lekidis † , Rasoul Behravesh ‡ , Avi Weit ⇤ ,
Pietro Giardina § , Vasileios Theodorou † , Cristina E. Costa ‡ , Katherine Barabash ⇤
⇤ † ‡ §
IBM, Haifa, Israel, Intracom Telecom, Greece, FBK, Trento, Italy, Nextworks, Pisa, Italy
email: {davidbr,weit,kathy}@il.ibm.com, {alekidis, theovas}@intracom-telecom.com,
{rbehravesh, ccosta}@fbk.eu, {p.giardina}@nextworks.it

Abstract—Network slicing is an essential 5G innovation have focused mostly on static single CSP slicing use cases,
whereby the network is partitioned into logical segments, so that where the slice resources and services belong to one domain,
Communication Service Providers (CSPs) can offer differentiated and their usage changes slowly in time once the slice is created.
services for verticals and use cases. In many 5G use cases,
network requirements vary over time and CSPs must dynamically However, some more complex use cases are emerging. These
adapt network slices to satisfy the contractual network slice QoS, require dynamic network slices, capable of following QoS
cooperating and using each others’ resources, e.g. when resources requirements variation and adapting the number of services and
of a single CSP are not sufficient or suitable to maintain all it’s resources employed to the actual demand. In such use cases,
current SLAs. While this need for dynamic cross-CSP cooperation services may span more than one CSP’s domain to properly
is widely recognized, realization of this need is not yet possible due
to gaps both in business processes and in technical capabilities. address demanding requirements, thus motivating the creation
In this paper, we present a 5GZORRO approach to dynamic of dynamic cross-domain slices.
cross-CSP slice scaling. Our approach both enables CSPs to Some system-oriented aspects of dynamic cross-domain 5G
collaborate, providing security and trust with smart multi-party
contracts, and facilitates thus achieved collaboration to enable
network slices have been recently addressed in EU H2020
resource sharing across multiple administrative domains, either projects [3–5]. In [6], an architecture that addresses technical
during slice establishment or when already existing slice needs to aspects of multi-POP slice establishment are addressed. Still,
expand or shrink. Our approach allows automating both business this area is in its infancy and present solutions can not easily
and technical processes involved in dynamic lifecycle management address some recent highly dynamic use cases, such as health-
of cross-CSP network slices, following ETSI’s Zero-Touch Net-
work and Service Management (ZSM) closed-loop architecture,
care and vehicular scenarios [7].
and relying on resource-sharing Marketplace, Distributed Ledger The 5GZORRO project [8] engages in a comprehensive study
(DL), and Operational Data Lake. We show how this approach on facilitating CSP collaboration to enable cross-CSP dynamic
is realized in truly Cloud Naive way, with Kubernetes as both slicing. 5GZORRO uses distributed Artificial Intelligence (AI)
business and technical cross-domain orchestrator. We then show-
case applicability of the proposed solution for dynamic scaling of to implement cognitive network orchestration and management
Content Delivery Network (CDN) service. with minimal manual intervention (zero-touch automation).
Index Terms—5G, Network Slicing, NFV, ETSI ZSM DLTs are adopted to implement flexible and efficient distributed
security and trust across the various parties involved in a 5G
I. I NTRODUCTION end-to-end service chain. Cloud-native orchestration at the busi-
5G telecommunication networks introduce drastically im- ness and technical level is employed to drive inter-operation of
proved Key Performance Indicators (KPIs) in terms of band- these technologies in the context of the business and technical
width, latency, reliability, availability, and device density. While workflows for dynamic cross-domain slice scaling.
this opens up possibilities for numerous novel applications, Collaboration between CSPs is an important principle that
the “one-size-fits-all” approach is highly inefficient due to allows mutually beneficial capacity sharing to mitigate resource
the diverse QoS requirements of applications. The need to scarcity during workload peaks and to extend services footprint.
simultaneously and cost-effectively address different use cases Traditionally, CSPs engage in manually-managed long-term
has motivated the use of novel technological solutions such peering agreements, suitable for inter-CSP collaboration for
as SDN and NFV to virtualize the physical 5G network into static cross-domain slicing, but inadequate for highly dynamic
multiple logical networks having different resource allocations cross-domain end-to-end network slicing scenarios of 5G and
and providing different QoS levels [1]. This approach is termed beyond. This is because static peering agreements result either
network slicing and offers isolation and differentiated QoS in over-provisioning or under-provisioning and do not reflect
guarantees to network silces [2]. the actual dynamics of the network usage by applications [9].
Network slicing enables multiple heterogeneous and service- The key challenges that hinder a more dynamic collaboration
specific logical networks to share a physical substrate network across CSPs are the lack of: (i) ubiquitous standards for dy-
that spans the Radio Access, the transport, and the core namic allocation and configuration of RAN, transport, and core
network. Until recently, industrial and research communities resources, standard mechanisms for ad hoc trust establishment,
978-1-6654-0522-5/21/$31.00 ©2021 IEEE and (iii) protocols for seamless inter-domain collaboration.
With the recent advancement of open standards, such as O-
RAN [10], Operator Platform Concept [11] and open source
projects like Free 5GC [12], traditional vertically integrated
telecommunication industry gradually opens up and creates an
ecosystem of standards abiding implementations that enable
better runtime collaboration across CSPs. In this work, we
focus on another aspect of inter-CSP collaboration, related to
lack of standard for coordination and orchestration of busi-
ness processes across domains, a major obstacle for efficient
resource sharing across CSPs. So far, this topic has not re-
ceived sufficient attention and we intend to fill this void by
presenting a position and making the first step towards a more
comprehensive study.
The basic design principles that we put forward include: (i) Fig. 1: ETSI ZSM closed-loop architecture [20]
autonomy of CSPs in their business and technical processes.
This includes freedom to use any NFV Management and approach has been widely applied to infrastructure orches-
Orchestration (MANO) framework and any NFV Infrastructure tration. In 5GZORRO, we follow this approach and create
(NFVI) platform as long as QoS of a slice network service domain specific Custom Resources (CR) and k8s operators
are satisfied; (ii) non-intrusive extensions to the existing NFV that continuously maintain CRs’ desired states. In addition,
MANO so that these existing frameworks are used as actuators we create custom Argo [15] workflows for 5GZORRO cross-
to the higher level orchestration mechanisms; (iii) preserve domain inter-CSP orchestration, building upon our previous
slice isolation so that dynamic cross-domain slice scaling does work reported in [16].
not disrupt or interfere with the QoS of slices and services For automatic translation from high-level business intent into
already provided in the involved domains; (iv) clean separation lower-level technical intents understood by k8s, we propose
of concerns between the technical and business aspects of leveraging Machine Learning (ML) algorithms such as Nat-
orchestration; (v) cloud-native declarative orchestration and ural Language Processing (NLP) [17]. NLP algorithms can
use of Kubernetes as the cross-domain control plane for scal- be trained to receive input in form of textual description of
able, portable, and reproducible business flow orchestration; the desired service characteristics and to produce a domain
(vi) rapid innovation orientation allowing to encompass novel specific encoding corresponding to the original intent. This
cross-CSP business models. process is step-wise: first, the intents are pre-processed, then
The rest of this paper is organized as follows: Sec- keywords are extracted and translated into meaningful actions
tion II introduces 5GZORRO slice scaling mechanisms; Sec- and, finally, aggregated and validated for lack of conflicts.
tion III presents workflow-based cross-domain slicing archi- Current NLP algorithms have limitations [18] and need to be
tecture along with relevant components; Section IV highlights further improved to be effectively used in 5G orchestration.
optimization-based slice scaling workflow; Section V gives B. Slice scaling overview
illustrative example with 5GZORRO vCDN use case; Section
VI concludes the paper and outlines future work. Our goal is to enable horizontal scaling of network slices.
For scaling-out, a subset of network service instances can
II. BACKGROUND be replicated and added to the network slice to handle the
increased workload, while for scaling-in unneeded network
A. Intent-based orchestration
service instances can be decommissioned to avoid unnecessary
Intent-based mechanisms are defined as means of automation costs. Horizontal scaling, that can take place over the infras-
in 5G networks, where each network action is automatically tructural resources of a single domain, as well be extended to
derived from the business needs instead of adhering to technical multi-domain setups, calling for inter-CSP collaboration and
network specifications. Three types of intent mechanisms are cross-domain resource discovery mechanisms [19]. We also
defined according to ETSI ZSM: (i) intent-based Application note that horizontal scaling actions have to be accompanied
Programming Interface (API), which is an interface for users to by automatic adjustment of additional resources, e.g., load
declare their intent, (ii) intent-based networking, which trans- balancing mechanisms.
lates the intent into a concrete network setup and (iii) intent- Our work goes beyond state-of-the-art through proposing a
based orchestration, which allows to control and orchestrate business-level orchestration approach harmonised with techni-
resources and network services dynamically. cal orchestration practices and consistent with the ETSI ZSM
A cloud-native approach to intent-driven management is [20] closed-loop architecture (Fig. 1).
embodied by Kubernetes(k8s) [13]. It orchestrates applications
though reconciliation loops based on controllers [14] and allows III. 5GZORRO A PPROACH
developing special mechanisms called K8s operators to define In this section, we first formulate the dynamic cross-domain
and manage custom resources. Being highly extensible, this slice scaling problem (Subsection III-A). Next, in Subsec-
Sixth, the complex cross-ISP interactions should execute
in reliable repeatable manner which calls for codification and
standardisation of the cross-domain process, so that they can be
driven by the automated orchestration. However, cross-domain
collaboration on slice scaling is not standardised yet and fast
innovation is required to address business aspects involved.
B. 5GZORRO Architecture
Fig. 2 outlines the 5GZORRO architecture as four groups of
functionalities. Detailed explanation of every functionality is
beyond the scope of this paper and is covered in [22]. We
focus only on the most relevant components, necessary for
understanding how we address the cross-domain slice scaling
challenges outlined in the previous subsection.
Fig. 2: A high-level 5GZORRO architecture Resource/Service Trading Functions (address the first and
third challenges): provide a Marketplace functionality pow-
tion III-B, we briefly outline the 5GZORRO approach and high- ered by DLT, including Smart Contracts, to enable trading of
light the importance of orchestration. In Section III-C, we focus 5G resources with no human intervention. Resource offers put
on the proposed orchestration and management mechanisms. on Marketplace follow [23] and include descriptive parameters,
such as cost, service area, Service Level Agreement (SLA), and
A. Cross-Domain Slice Scaling Challenges
observability. The marketplace resources can be discovered
Dynamic cross-domain slice scaling poses a number of through the Smart Resource Discovery (SRSD) component that
challenges. First, while a CSP is aware of its own resources offers two interfaces: (a) simple catalog interface and (b) a more
and network services inventory, network topology, and service sophisticated search interface that returns resource selection
areas, it does not have the same information about another CSP. where resources are already scored by some criteria.
Second, multiple trade-off optimisation might be involved Security and Trust Functions (address the third challenge):
in resource selection optimisation (e.g., cost-efficiency, energy help to ensure the authorization, authentication and integrity of
efficiency, etc.). Typically, these problems are NP-hard. Given all communications and transactions within the platform, and
a large number of potential resource choices, high frequency of provide distributed identity, and secrets management function-
decision making, on-line nature of the problem (slice scaling alities for the 5GZorro platform participants as well as reliable
requests are not known in advance), heuristics are required to trust and security scores for Marketplace resources.
find good sub-optimal solutions fast enough. Also an overhead Analytics and Intelligence Functions (address the second
of configuration should be taken into account. and fourth challenges): provide functionalities in the Orient
Third, when a CSP procures resources from another one, and Decide blocks of the ETSI closed-loop architecture. In
this transaction should assure non-repudiation, avoid resource 5GZORRO, these functions can be deployed by CSPs internally
double booking, and deny unilateral retractions of the resources (a usual use case) as well as on top of a logically centralised
by one of the leasing side. Given a dynamic nature of the Data Lake. A typical example of this functionality is slice
problem and a need to avoid subjective judgement, a human optimisation. Another typical example is SLA compliance
operator cannot be placed in the loop and these transactions prediction [24]. Each technical resource/service owner CSP
should occur automatically, but subject to the terms and policies pushes monitoring data into the Data Lake and the legal
set by the human agents in the respective CSPs. owner of the resources can execute analytic pipelines that
Fourth, once a slice is scaled into another domain, monitor- infer predictive models on the incoming data to forecast SLA
ing information should be exchanged and compliance to Service violations. Likewise, analytic pipelines can be run on the Data
Level Agreements (SLA) should be monitored and enforced. Lake to [re-]train these models.
This poses the following questions: at which level the technical Zero-Touch Management and Orchestration Functions (ad-
owner of a resource should share monitoring data with the legal dress aspects of the sixth challenge): comprises two types of
owner of a resource? What mechanisms should be deployed to management and orchestration functionalities: (a) technical and
facilitate this sharing in a secure and trusted way? Another (b) business-level. The technical orchestration and management
important issue is how to factor in the cost of monitoring (e.g., functionality (NS and Service Orchestrator (NSSO)) adheres to
hauling monitoring data from the technical to legal owner?). the MANO standards, such as those by ETSI NFV MANO. An
Fifth, the slice scaling is driven by requirements of appli- innovative aspect of the technical orchestration in 5GZORRO
cations, e.g., those running in Multi-Access Edge Computing is its ability to horizontally inter-operate with other MANO
(MEC) Platforms (MEP) as suggested in [21]. However, MEC engines (a feature that we leverage to facilitate technical
applications cannot directly influence network slices so there is orchestration across domains). The higher-level orchestrator
a need to mediate between the high level dynamic application (Intelligent Slice and Service Manager (ISSM)) manages and
intents, the MEC orchestrators, and the slice orchestrators. orchestrates components that are outside of the scope of MANO
Event Mediator. Intents can be submitted to ISSM through three
different channels: (a) a human operator submits a high level
intent in natural language and it is transformed by NLP into
a lower level intent conforming to e.g., GSMA NEST [2], (b)
MEC application control plane, and (c) continuous optimisation
executed by ISSM-O.
An intent reception causes a codified business flow to execute
by ISSM-WFM. The workflow that reaches the appropriate
5GZORRO components to automate their interactions and
facilitate the slice scaling mechanisms. During the workflow
execution, 5GZORRO components are triggered in order to
Fig. 3: Intelligent Slice and Service Manager architecture orchestrate their business process. One such component is
the Smart Resource and Service Discovery (SRSD), which
(Marketplace, DLT, Data Lake, SRSD, Security and Trust, etc.), allows to discover relevant resources and then purchase them
which are essential to tackle the challenges of the 5G slicing to form or extend a NS. SRSD follows an NLP approach
cross-CSP scale out as described in Section III-A. Zero-Touch to translate intents originating from an API into concrete
Management and Orchestration is guided by Machine Learning requests on resources that are also ranked according to their
(ML) technologies to ensure automation of orchestration and relevance for the given intent. Following the ranking process,
management in NSs and the services they are handling. the best resource candidates are selected, which allows ISSM
Intelligent Slice and Service Manager (ISSM), described in to navigate to a Marketplace Portal for purchasing them.
the next subsection, ties up interactions among the 5GZORRO The workflows themselves are declaratively defined as Ku-
components into workflows, e.g. for scalable slice scaling. bernetes CRs in YAML and represent events and dependencies
C. Intelligent Slice and Service Manager that are triggered by these events.
As we introduced in the previous subsections, the challenges ISSM-WFM allows executing multiple workflows asyn-
posed by cross-domain slice scaling call for new platform chronously and handles error conditions. The progress of the
components. The ETSI ZSM reference architecture specifies workflows can be examined at any time. ISSM-O monitors
intra-domain and cross-domain Integration Fabric functionality performance of slices as reflected on the Data Lake as well as
and E2E Service Orchestration functionality [20]. The Inte- the Marketplace resource offers. ISSM-O continously proposes
gration Fabric is a special management function that allows slice optimization intents that, in particular, include slice scal-
cross-domain service exposure and composition. It enables ing to ensure the the QoS of each slice (described in Section
registration, discovery and invocation of management services IV). These events trigger ISSM-WFM workflows to execute.
and the communication between management functions. Likewise, a MEC application control plane might submit an
The E2E Service Orchestration functionality of ETSI ZSM intent to ensure its Quality of Experience (QoE). This intent is
provides services related to E2E lifecycle management of a transformed into a lower level intent by ISSM-O. In particular,
service within and across domains. However, it does not offer this might include slice cross-domain scaling, which will trigger
details on which component should drive the high-level inter- a corresponding ISSM-WFM workflow.
actions enabled by Integration Fabric and E2E Orchestration. On the technical side, the slice extension is performed
The cross-domain slice scaling management entails a high through the use of the NS and Service Orchestrator module
level inter-operation workflow involving the aforementioned (NSSO), which represents the actuation part in the slice and
components of the platform. It is facilitated in 5GZORRO by service orchestration stack, responsible for the vertical and
ISSM, which focuses on the automation of the management network service deployment across different administrative
of secure cross-domain slices and of the services within them domains, under the responsibility of two main internal modules:
(including optimisation). Hence, its components can be placed Vertical and Network Service Management Function (VSMF
in the Decide and Act blocks of the ETSI ZSM closed-loop and NSMF). The NSSO abstracts the concept of vertical service
architecture (Fig. 1). by using custom templates called Vertical Service Blueprints
In our approach, ISSM is a generic cloud-native orchestrator (VSB) representing a class of Vertical Services. Once a VSB
following the design principles outlined in Section I and relying is filled with specific values it becomes a Vertical Service
on Kubernetes as this is a de facto enabler for cloud-native Descriptor (VSD) which actually represents the Vertical Service
software. As such, it can be programmed to execute any high and its characteristics. The set of parameters that can be
level flows, not only the slice scaling related ones. specified inside a VSB includes the list of network func-
Fig. 3 shows the ISSM’s high-level architecture. ISSM tions and vertical applications to be deployed (called “atomic-
comprises three components: (1) Workflow Manager (ISSM- components”), along with all the information required for the
WFM), (2) Optimiser (ISSM-O), and (3) MEC orchestrator orchestration in terms of dimensioning (e.g., bandwidth, num-
(ISSM-MEC). These components are interconnected via Event ber of users/sessions, etc). Various aspects of the connectivity
Bus. The events are forwarded to the right subscribers via can be also set, in terms of service type (e.g., L2/L3 VPN),
QoS, Internal and External interconnections (i.e. intra or cross-
domain slice/service connections). In 5GZORRO, the ISSM
creates the VSD by specifying in the selected VSB the proper
parameter values derived by the results of the process of smart
resources selection, and requests its instantiation to the NSSO.
At instantiation time, the VSD is translated into a set of NS
Templates (NST), partially aligned with 3GPP specification
[25], whose lifecycle is managed by the NSMF by direct
interaction with the NFVOs. VSMF and NSMF are NSSO
modules running separately under a one-to-many relationship
and interacting through the REST interface exposed by the
NSMF, where each NSMF can support several NFVOs. The
This design enables high flexibility in the deployment of the
NSSO and in the management of different services and slices
across different administrative domains. A VSMF peer-to- Fig. 4: A slice establishment workflow
peer mechanism is supported, allowing the deployment of a
federation of NSSOs. B. Workflow Automation
With the proposed method we aim at resolving of two main
IV. S LICE O PTIMISATION scenarios: 1) the automation of the slice establishment across
multiple CSPs and 2) the slice scaling in the case that an
Different from resource optimisation in a single domain,
existing CSP slice needs to be extended with the resources from
in which the CSP has complete information about the net-
another CSP when its own are saturated. For both challenges,
work topology and state of resources, the problem becomes
the ISSM provides a concrete workflow that is automated using
complicated when it comes to optimisation across domains,
Argo. Here we focus on the first scenario, while the second is
because CSPs do not access to the topology and resources
depicted in Fig. 6.
information of other CSPs. In this regard, the optimisation
The ISSM workflow for slice establishment involves two
should be performed at different levels, namely intra- and inter-
CSPs (CSP A and CSP B) and consists of the following steps
domains.
illustrated in Fig. 4:
1) The service owner starts a business process by submitting
A. Optimisation technologies
an intent on slice profile to ISSM and the latter starts
We adopt an approach presented in [26]. Although that work executing an explicitly codified business flow.
performs slice placement in different technological domains, 2) The ISSM-WFM requests the ISSM-O to translate the
namely RAN, core, and cloud under the control of single provided slice blueprint into intent-based resource requests
CSP domain, it can be adapted to the scenario of multi-CSP for the SRSD.
optimisation by carefully defining the properties of available 3) The ISSM-WFM calls the SRSD to perform contextual-
resources in the 5GZORRO Marketplace. ized queries for available resources.
The NSSO slice optimisation is a two-step process that 4) The SRSD returns the obtained resources to the ISSM-O
leverages two specialised modules, called Translator and Arbi- for further optimization.
trator. These modules implement the decision logic in terms of 5) The ISSM-WFM uses DLT to acquire resources.
resource selection and optimisation during the instantiation and 6) The ISSM-WFM links the retrieved resources with the
actuation processes. The NSSO implements its own Arbitrator slice blueprints of the NSSO for each CSP using Translator
and Translator that base their decisions on a set of policies and Arbitrator.
and rules. ML-based optimisation algorithms implemented by 7) The slice subnets are formed by the NSSO.
external modules are also supported. The Translator module 8) The slice subnets are activated and glued together through
associates to a VSD a set of proper NSTs that will contain NSSO orchestrating a tunnel creation via SDN.
proper resource specifications to meet the requirements of At the end of this process, the formed NS includes VNFs
the Vertical Service as set of NSs and Vertical Applications. containing applications from different CSPs and 5G infrastruc-
NSTs may contain, as result of the translation process, RAN ture resources of both operators. In example shown in Fig. 4,
specification, Network Service Descriptors, characterized by NS includes a MEC [21] platform, RAN resources and a UPF.
a properly selected deployment flavour, and QoS constraints The VNFs, along with the infrastructure resources, are glued
as specification for the building of the NS(s) associated to together through virtual links. Upon NS establishment, ISSM-
the Vertical Service. The Arbitrator checks the requirements WFM can instantiate an SLA breach prediction pipeline in the
expressed in the various NSTs and decides on the most ap- 5GZORRO Operational Datalake and CSPs can start pushing
propriate way to deploy the required services and slices while annotated monitoring data for this NS’s resources through their
satisfying feasibility constraints. own data ingestion pipelines
1 "id": 127218,
2 "resourcename": "virtual node",
3 "resourceType": "edge",
4 "location": "Madrid",
5 "category": "vCDN",
6 "transactionID": 145678,
7 "resourceVirtualCapabilities": [
8 {
9 "nodeId": "51787e4a-1783-414b-82f9-95cd85231724",
10 "isMaster": 0,
11 "type": "storage",
12 "virtualCapabilities": {
13 "NoOfCPUcores": {
14 "virtualCapKey": "SSD",
15 "virtualCapValue": 100,
Fig. 5: ISSM GUI (Argo based) 16 "virtualCapUnit": "GB"
17 } } } ]
V. CDN USE - CASE IN 5GZORRO Listing 1: SRSD resource representation snippet
In this section, we demonstrate a scenario that leverages the
method proposed in Section III to mitigate the challenge of quires the resources from the 5GZORRO Marketplace. When
resource saturation of a CSP, who is leasing a NS instance from all resources are acquired, ISSM-WFM instructs the NSSO to
another CSP to provide a vCDN service for end users. The slice create the slice subnet for the vCDN service. A snippet of such
embeds performance guarantees related to e.g., throughput and subnet is illustrated in Listing 2.
latency in certain service areas, based on a profiling of the
service workload. Large-scale social events or very densely 1 "vsBlueprint": {
2 "name": "VideoStreaming",
populated areas demonstrate workloads significantly above the 3 "description": "Entertainment vCDN service",
predefined workload profile (e.g., number of ongoing video 4 "nsdIdentifier": "vCDN_NSD_v01",
5 "nsdName": "vCDN Network Service",
streams), rendering default resource allocation inadequate. To 6 "parameterId": "bandwidth" [
provide improved QoE to the end users, CSPs can rely on edge 7 "minValue": "20",
8 "maxValue": "40" ]
computing technologies, such as MEC platforms [21]. Spikes 9 virtualLinkDesc: [
in demand may exhaust the compute capacity, e.g. for video 10 {
11 connectivityType: {
transcoding or object recognition, and impact the slice QoS. 12 layerProtocol: "IPV4"
Our vCDN scaling use case consists of several steps. First, 13 },
14 qos: {
the slice is established using the workflow depicted in Fig. 4. 15 latency: 10ms,
The workflow specification is a YAML Argo CR that specifies 16 packetLossRatio: 0.001%
17 } } ], }
dependency on instantiating a GSMA-based [2] slice template.
We omit the YAML specification due to the lack of space and Listing 2: NSSO slice subnet
show a snippet of the Argo Graphical User Interface (GUI)
executing the workflow in Fig. 5. Next, a custom intent towards This slice subnet includes information on QoS requirements
the SRSD is created with phrases at different abstraction levels, of the vCDN service (bandwidth) as well as on the individual
such as: ”Edge-based near Madrid with more than 60 GB virtual links (virtualLinkDf ). Additionally, the connectivity
RAM” or ”minimal edge storage nearby”. Then, intent phrases protocol that is used is also defined (layerProtocol). This slice
are translated using NLP techniques, e.g. cossine or jackard subnet is then provided to ISSM, in order to stitch it with the
similarity [17], into a concrete query to the SRSD resource second CSP that has usable and spare resources.
database, then results are ranked uses ranking algorithms.
The highest ranked resource returned to ISSM- based on the VI. C ONCLUSION
given intent, is depicted in Listing 1. The resource representa- This paper presented a novel approach to dynamic NS
tion is following the TM Forum 634 [23], which is stardardizing scaling through dynamic collaboration between multiple CSPs
the definition of resource offers. supported by DLT transactions. The collaboration on the busi-
In this listing, the characteristics (resourceType) and capabil- ness process side allows resource sharing across different 5G
ities (resourceVirtualCapabilities) of a resource are provided. infrastructure segments and on the technical process side, slice
Additionally, there is the transactionID that refers to the DLT stitching from different CSPs or the extension of already
transaction identifier that CSPs use to annotate the data so that formed slices for ensuring QoS requirements. Furthermore,
it belongs to this transactional context. Moreover, within these both processes are automated using ML techniques inspired
capabilities the type of resource is included (storage type in by ETSI’s ZSM closed-loop architecture.
Listing 1). The format for this resource type (virtualCapKey) In this work, we applied this approach to 5GZORRO vCDN
as well as its value (virtualCapValue) and unit (virtualCapUnit) use-case that focuses on providing QoS in large-scale social
are also part of the returned response. events as well as in densely populated areas. In this use-case,
Next, ISSM-O can use the provided representation to per- the increased demands lead to a saturation of resources in the
form further optimizations (Section IV) and ISSM-WFM ac- CSP 5G infrastructure, and hence the slice is extended with
[7] Kukkalli et al., “Evaluation of Multi-operator dynamic 5G
Network Slicing for Vehicular Emergency Scenarios,” in
2020 IFIP Networking Conference (Networking), 2020.
[8] Gino Carrozzo et. al., “AI-driven Zero-touch Operations,
Security and Trust in Multi-operator 5G Networks: a
Conceptual Architecture,” in EuCNC 2020, 2020.
[9] T. Subramanya and R. Riggio, “Centralized and Feder-
ated Learning for Predictive VNF Autoscaling in Multi-
Domain 5G Networks and Beyond,” IEEE Tran. on Net.
and Serv. Man., 2021.
[10] “O-RAN Alliance,” 2021. [Online]. Available:
https://www.o-ran.org/
Fig. 6: Saturation of CSP resources and corresponding slice [11] “Operator Platform Concept, phase 1: Edge cloud com-
scaling actions puting.”
[12] “Free 5G Core,” 2021. [Online]. Available:
additional resource from other CSPs. Our approach has been https://www.free5gc.org/
validated in a development environment. [13] Cloud-Native Computing Foundation (CNCF), “CNCF
The business level workflows orchestration have been im- Cloud Native Definition v1.0,” 2018.
plemented using Kubernetes native workflow management. We [14] “Kubernetes Controllers,” 2021. [Online]. Available:
believe that Kubernetes can play a threefold role: (a) VIM, https://kubernetes.io/docs/concepts/architecture/controller/
(b) MANO, and (c) business level orchestrator. The first role is [15] “Argo Project: Open Source Kubernetes Native
widely utilised. The second role was previously reported in [16] Workflows, Events, CI and CD,” 2021. [Online].
and the third one is explored in this work. Available: https://argoproj.github.io/
As a part of our future work, we plan to evaluate our [16] D. Breitgand, “Using Argo and Knative to Orchestrate
architecture in a larger testbed and with additional use case, Media-intensive Services in 5G Edge,” in KubeCon and
to quantify the benefits that are accrued for CSP through inter- CloudNativeCon Europe 2020.
CSP slice scaling in use cases studied in [27]. [17] H. Schütze and et. al, Introduction to information re-
trieval. Cambridge University Press Cambridge, 2008.
ACKNOWLEDGMENTS [18] C. Benzaid and T. Taleb, “Ai-driven zero touch network
and service management in 5g and beyond: Challenges
This work has received funding from the European Union’s and research directions,” IEEE Net., 2020.
Horizon 2020 research and innovation programme under grant [19] Adamuz-Hinojosa et al., “Automated network service
agreement No 871533 (5GZORRO). scaling in nfv: Concepts, mechanisms and scaling work-
flow,” IEEE Comm. Mag., 2018.
R EFERENCES [20] ETSI, “Zero-touch network and Service Management
(ZSM); Reference Architecture – ETSI GS ZSM 002
[1] X. Li et al., “Network Slicing for 5G: Challenges and V1.1.1 (2019-08),” Jul 2019.
Opportunities,” IEEE Internet Computing, 2017. [21] ETSI, “GR MEC 017: Mobile Edge Computing
[2] GSMANG.116, “Generic Network Slice Template Version (MEC);Deployment of Mobile Edge Computing in an
3.0,” 2020. NFV environment,” 2018.
[3] “SliceNet: End-to-End Cognitive Network Slicing and [22] 5GZORRO, “Design of the 5GZORRO Platform for Se-
Slice Management Framework in Virtualised Multi- curity & Trust,” 2021.
Domain, Multi-Tenant 5G Networks,” 2020. [Online]. [23] TM Forum, “TMF634 Resource Catalog API User Guide
Available: https://slicenet.eu/ v4.1.0,” 2020.
[4] “5G-MONARCH: 5G Mobile Network Architecture [24] V. Theodorou et al., “Blockchain-based Zero Touch Ser-
for diverse services, use cases, and applications in vice Assurance in Cross-domain Network Slicing,” To
5G and beyond,” 2019. [Online]. Available: https://5g- appear in the proceedings of EUCNC/6G summit 2021.
monarch.eu/ [25] 3GPP, “Management and orchestration; 5G Network Re-
[5] “Novel Enablers for Cloud Slicing (NECOS),” 2019. [On- source Model (NRM); Stage 2 and stage 3,” Mar 2020.
line]. Available: http://www.h2020-necos.eu/motivation- [26] Behravesh, Rasoul et. al, “Joint User Association and
and-vision/ VNF Placement for Latency Sensitive Applications in 5G
[6] P. Valsamas, P. Papadimitriou, I. Sakellariou, S. Petridou, Networks,” in Proc. of IEEE CloudNet, 2019.
L. Mamatas, S. Clayman, and F. Tusa, “Multi-PoP Net- [27] Sayadi, Bessem et al., “5G-PPP Software Network Work-
work Slice Deployment: A Feasibility Study,” in 2019 ing Group Cloud-Native and Verticals Services,” 08 2019.
IEEE 8th International Conference on Cloud Networking
(CloudNet), 2019.

View publication stats

You might also like