PreprintPDF Available

Rapid prototyping and performance evaluation of MEC-based applications

Authors:
Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

Multi-access Edge Computing (MEC) will enable context-aware services for users of mobile 4G/5G networks. MEC application developers need tools to aid the design and the performance evaluation of their apps. During the early stages of deployment, they should be able to evaluate the performance impact of design choices - e.g., what round-trip delay can be expected due to the interplay of computation, communication and service consumption. When a prototype of the app exists, it needs to be tested it live, under controllable conditions, to measure key performance indicators. In this paper, we present an open-source framework that allows developers to do all the above. Our framework is based on Simu5G, the OMNeT++-based simulator of 5G (NewRadio) and 4G (LTE) mobile networks. It includes models of MEC entities (i.e., MEC orchestrator, MEC host, etc.) and provides a standard-compliant RESTful interface towards application endpoints. Moreover, it can interface with external applications, and can also run in real time. Therefore, one can use it as a cradle to run a MEC app live, having underneath both 4G/5G data packet transport and MEC services based on information generated by the underlying emulated radio access network. We describe our framework and present a use-case of an emulated MEC-enabled 5G scenario.
Content may be subject to copyright.
1
Rapid prototyping and performance evaluation of MEC-based applications
Alessandro Noferi1, Giovanni Nardini2, Giovanni Stea2, Antonio Virdis2
Dipartimento di Ingegneria dell’Informazione, University of Pisa
Largo L.Lazzarino, 1, 56122, Pisa, Italy
1alessandro.noferi@phd.unipi.it, 2{name.surname}@unipi.it
Abstract Multi-access Edge Computing (MEC) will enable context-aware services for users of mobile 4G/5G
networks. MEC application developers need tools to aid the design and the performance evaluation of their apps. During
the early stages of deployment, they should be able to evaluate the performance impact of design choices - e.g., what
round-trip delay can be expected due to the interplay of computation, communication and service consumption. When a
prototype of the app exists, it needs to be tested it live, under controllable conditions, to measure key performance
indicators. In this paper, we present an open-source framework that allows developers to do all the above. Our framework
is based on Simu5G, the OMNeT++-based simulator of 5G (NewRadio) and 4G (LTE) mobile networks. It includes
models of MEC entities (i.e., MEC orchestrator, MEC host, etc.) and provides a standard-compliant RESTful interface
towards application endpoints. Moreover, it can interface with external applications, and can also run in real time.
Therefore, one can use it as a cradle to run a MEC app live, having underneath both 4G/5G data packet transport and
MEC services based on information generated by the underlying emulated radio access network. We describe our
framework and present a use-case of an emulated MEC-enabled 5G scenario.
Keywords Simulation, Emulation, MEC, Simu5G, Real-time, prototyping
1. Introduction
Multi-access Edge Computing (MEC) will deliver cloud-computing capabilities at the edge of the network. Besides
providing a smaller and more predictable latency, this will enable context-awareness, capitalizing network information
such as radio access conditions, user location, etc., which would be difficult to know, if not impossible, in a cloud-based
application. This is achieved by having MEC communicate with the access network, via the so-called MEC services. These
are services exported by a MEC platform connected to the access technology, accessible via a RESTful interface, that a
MEC application (MEC app, henceforth) can query to acquire the user context. They provide information on the user (e.g.,
its own radio conditions or location) as well as on the access itself (e.g., what is the current network load or number of
users), thus allowing one to create advanced user services. Examples of these user services already discussed in the
context of the European Telecommunications Standards Institute (ETSI), who is standardizing MEC are: user QoS
prediction based on current radio conditions, location and movement pattern; vehicular alerts, e.g. a car approaching a
slippery patch of tarmac or robot swarm coordination in a factory.
While, as its very name suggests, MEC is access-technology agnostic, it is quite clear that its interplay with mobile
networks, namely 4G and 5G cellular ones today and 6G tomorrow will be prominent in the future. This is because
ubiquitous, regulated, reliable and secure wireless access is a pre-requisite for marketable user services. In this scenario,
MEC services are provided by the underlying 4G/5G radio access, leveraging the information base available at the base
2
stations (eNB in LTE, gNB in New Radio) and in the various network entities involved in the control and management
plane. In this paper we refer to this scenario, without explicitly repeating it henceforth.
By making the hosting infrastructure open under controlled conditions MEC enables the role of MEC app
developer. This will not necessarily be related to either the network operator or the MEC infrastructure provider. MEC
apps developed by this player will be hosted in the MEC infrastructure, and they will communicate with the user (on
behalf of whom they are run) through the cellular network run by the operator. More to the point, the MEC app will also
interact with the cellular network by consuming MEC services provided by it. It is foreseen that the growth of the
community of MEC app developers will be the driving force for a widespread MEC diffusion.
In this scenario, it becomes particularly important for MEC app developers to be able to test their applications for
functionality and performance in credible and controllable settings. By credible, we mean taking into account: i) radio
resource contention on the data plane of the underlying cellular network; ii) resource contention at the MEC host, given
that MEC apps run as virtual machines or containers on a shared computing infrastructure, and iii) contention for access to
MEC services (e.g., queueing delays), due to concurrent service requests. These settings must be fully controllable, e.g.
allow one to test the impact of radio congestion, poor channel, MEC platform congestion, etc., especially if the user
services are time-critical (as are some of those mentioned at the beginning of this section) or come with service level
agreement guarantees. Functional and performance testing of MEC apps is required at different stages in the development
cycle: early on, one may want to understand the performance implications of alternative designs, e.g., the ensuing
communication patterns, before committing to a given solution. Later, one may want to test a prototype of its MEC app in
real time, measuring delays and possibly factoring in user interaction. Eventually, a developer may want to showcase its
own MEC app, completed and polished, in a live demo, e.g. for sales pitch to prospective financers. All this is made
difficult by the lack of suitable tools, and the fact that it is hardly possible to use a live 5G network. There are indeed tools
that simulate end-to-end communications on cellular networks see, e.g., [16][17], tools that simulate the MEC
environment see, e.g., [23] and survey [24], open-source environments to host MEC apps see, e.g. [28]. There are also
tools that allow one to test MEC service interfaces see, e.g., [41][42]. However, none of these can work in synergy,
hence while certainly helping a developer they do not allow the kind of support to prototyping discussed above.
In this paper we describe an open-source complete framework to integrate all the above functionalities. It is based on
Simu5G [6][7], an open-source simulation library for 5G (and 4G) cellular networks based on OMNeT++ [8]. Our
framework models the whole MEC infrastructure including the MEC orchestrator, the MEC host running MEC apps and
the MEC platform providing MEC services in an environment devised for end-to-end, application-oriented discrete-
event simulation. All the interfaces between the MEC environment and the application world are designed to be ETSI-
compliant. Moreover, our framework can interface with external application endpoints and carry their traffic, and can do
so in real time [2][3], acting as a network emulator [4][5]. A MEC app developer can thus exploit it as a cradle for
distributed MEC-based applications, which can be both tested for functional compliance and evaluated for performance,
maintaining full control over the experimental setup. For instance, a developer may want to test a UE app (the one residing
on the user device) with a stub of a MEC app (residing on the MEC host), which she can quickly develop as an OMNeT++
module within Simu5G. Alternatively, she may want to test a MEC app that consumes MEC services (e.g. the Radio
Network Interface Service RNIS [31], or the Location Service [32]), or find under what load conditions the response
time of said MEC services may affect the end-to-end application performance. In this case, the MEC app could be
developed within Simu5G, or outside it, e.g., hosted on a real MEC host such as Intel OpenNESS [28], consuming MEC
3
services provided by Simu5G. For instance, developers of an edge-assisted Virtual Reality application [10] may be
interested in assessing the Quality of Experience perceived by users when the application runs in a MEC system and
exploits context information from a realistic 5G network scenario.
Two comments are in order. The first one is that as far as we know there seems to be no tools comparable to the
one described in this paper. If you want to perform a live run of your MEC app in a 5G network, it seems that the only
alternative is that you avail yourself of a 5G network testbed and a MEC infrastructure connected to it, with MEC services
enabled, and run your app over them. This is expensive, time consuming and might still not allow you full control over
experimental conditions (say, to create congestion or radio impairments). Our second comment is that the work presented
in this paper can be expected to be useful also to the other players involved in the MEC chain of value as well. On one
hand, a cellular network operator will need to test the impact of MEC-generated traffic on its Radio Access Network data
plane, as well as the impact of MEC service requests on its network’s control/management plane. On the other hand, MEC
infrastructure providers may want to use these tools to plan for capacity, evaluate the performance impact of a growing
number of hosted MEC apps on the MEC platform, identify bottlenecks and enforce admission control, etc.. Moreover,
this work is currently being used to support the demonstration of federated learning of eXplainable Artificial Intelligence
(XAI) models within 6G flagship EU project Hexa-X [29].
We describe the modeling of the MEC subsystem within Simu5G and show how these models are built for scalability,
i.e., allow one to simulate congestion at the MEC platform efficiently, so that they still allow one to emulate in real time a
large-scale computation and communication scenario (hundreds of simultaneous MEC apps, complex 5G network with
tens of nodes and hundreds of users) on a desktop PC. Moreover, we describe an example of a MEC-based application
running live through our framework, showing that setting up a testbed is quite simple and requires inexpensive hardware
resources.
The rest of this paper is organized as follows: Section 2 provides background information on the technologies we use.
Section 3 and Section 4 present the architecture of our software framework, focusing on the model of the MEC
infrastructure and the MEC applications and services, respectively. Section 5 presents a use case. Section 6 reviews the
related work. Finally, Section 7 draws conclusions and highlights directions for future work.
2. Background
In this section, we provide background knowledge required to understand the design choices for our MEC prototyping
framework. More specifically, we present the capabilities of the Simu5G OMNeT++-based library, and we introduce the
reference architecture of ETSI MEC.
2.1. Overview of Simu5G
Simu5G [6][7] is the evolution of the well-known SimuLTE 4G network simulator [25] towards 5G NewRadio (NR)
access. It simulates the data plane of both the core and the radio access networks. Since it incorporates all SimuLTE’s
functionalities, it allows users to create legacy or mixed 4G/5G scenarios as well. Hereafter, we describe those entities and
functionalities of Simu5G that are more closely related to the scope of this paper. We refer the interested reader to [6][7]
for more details.
Simu5G is a model library for the OMNeT++ discrete-event simulation framework [8]. The latter allows analyzing any
kind of system in which there are entities communicating with each other. In OMNeT++ these entities are represented by
simple modules communicating via message exchange through connections among module gates. Module’s behavior is
4
coded in C++. Multiple modules can be composed to form a compound module. Henceforth, we will refer to modules to
denote either simple or compound OMNeT++ modules. Simu5G is interoperable with all the libraries based on
OMNeT++, such as INET [9] for TCP/IP-based network technologies, and Veins [11] for vehicular mobility. This allows
users to construct very complex scenarios straightforwardly, by importing and connecting existing libraries, with hardly
any extra line of code required. Simu5G, in fact, leverages several models and functionalities taken from INET itself.
Figure 1 shows the main modules of Simu5G.
As far as the core network (CN) is concerned, Simu5G defines a User Plane Function (UPF) or Packet GateWay
(PGW) and allows users to construct arbitrary CN topologies, where packet forwarding is based on the GPRS tunneling
protocol (GTP). For what regards radio access, Simu5G defines gNBs and UEs, which communicate using the New Radio
protocol stack at layer 2. gNBs can be connected to the CN directly, in the so-called StandAlone deployment, or operate in
an E-UTRA/NR Dual Connectivity (ENDC) deployment, where LTE and NR coexist. In this last configuration, the gNB
works as a secondary node for an LTE eNB, which acts as master node connected to the CN [12].
Both UEs and gNBs include a NR Network Interface Card (NIC), which models the NR protocol stack. With reference
to Figure 2, the NR NIC includes all the sublayers of NR (from the Packet Data Convergence Protocol to the Physical
Layer), with 3GPP-compliant behavior. As for the physical layer, Simu5G follows the approach already used by
SimuLTE, i.e. to model the effects of propagation on the wireless channel at the receiver, without modelling symbol
transmission and constellations. This allows us to compute the Signal-to-Interference-and-Noise Ratio (SINR) at receivers
correctly and to compute MAC-layer decoding probabilities accordingly, at a manageable complexity, and makes the
simulator more evolvable.
Simu5G simulates radio access on multiple carriers, in both Frequency- and Time-division duplexing (FDD, TDD).
Different carrier components can be configured with different FDD numerologies and different TDD slot formats.
Moreover, different carrier components can have different channel models and MAC-level schedulers. Simu5G
incorporates UE handover and network-controlled device-to-device communications, both one-to-one and one-to-many.
Simu5G can run in real time, thanks to OMNeT++ and INET functionalities. In fact, OMNeT++ allows events to be
scheduled by a real-time scheduler, which matches the simulated time to the wall-clock time at the beginning of a
simulation, and slows down event scheduling until the corresponding wall-clock time has expired. This is only possible, of
course, if simulated time would otherwise run faster than wall-clock time, something which depends on both the
hardware/software system that runs OMNeT++, and the scenario being simulated. More specifically, the larger the scale of
the simulation (e.g., the higher the number of UEs and gNBs), the less likely it is that this is possible. Parallel to this, INET
allows one to endow modules with external interfaces that exchange packets with the outside world. This means that an
Figure 1 - Main modules of the Simu5G model library
Figure 2 - Internal model of the NR NIC
5
external application can exchange packets with a module in a simulated network. Since simulation can also run in real
time, this allows one to perform real-time network emulation, transporting packets between external application endpoints.
Figure 3 shows a simple scenario wherein host A and C are physically connected to host B, which runs an emulated
scenario in Simu5G. Packets sent by host A appear in the emulation at the UE, and packets leaving the rightmost interface
of the router are sent to host C. Packets flows similarly in the other direction as well. Both the UE and the router must be
endowed with external interfaces for this to happen. Provided that routes are added to host B’s operating system, an
application on host A can thus reach any entity having an IP address in the emulated scenario (e.g., the UPF). Moreover, it
can reach host C, with its packets traversing the emulated scenario. More complex configurations can also be created, such
as having one remote host reachable using a public IP address.
Simu5G inherits the above capabilities being based on OMNeT++ and INET but is also explicitly designed for
scalability, i.e., to allow real-time emulation of relatively large scenarios on a desktop machine. As shown in [3], it offers
multiple models of entities (i.e., UEs and gNBs), which can be used for different purposes: foreground entities run the
entire protocol stack and can be connected to external interfaces. On the other hand, background entities run only those
functions that are necessary to create communication impairments: more specifically, background UEs and gNBs create
lifelike interference at the physical layer, and background UEs contend for access to resources at the MAC layer.
Background entities, however, have a reduced protocol stack and do not transmit packets over the air, hence cannot host
external interfaces. Background entities create the same impairment that complete entities would, at a much lower
overhead, which allows one to extend the scale of scenarios that can be emulated in real time. Work [3] shows that one can
emulate scenarios with tens of gNBs and several hundreds of UEs, where two endpoints exchange several Mbps’ worth of
traffic, all on a desktop PC.
2.2. ETSI MEC Architecture
MEC is a computing infrastructure that can run applications (MEC apps) as virtual machines or containers on behalf of
network users, interfacing with the access network (e.g., a cellular network). The MEC architecture has been standardized
by the Industry Specification Group of ETSI [34]. The above standard covers the functional entities of the MEC
architecture, as well as the procedures for a user to instantiate a MEC app. The data-plane exchange between a user and a
MEC app is not part of the standard. This section describes the entities that compose the MEC architecture, with a focus on
the ones implemented in the framework, which are highlighted in black in Figure 4. The MEC architecture includes a MEC
host level and a MEC system level. The former contains the virtualization infrastructure used to run the virtual machines
containing MEC apps. Within it, the MEC platform maintains a Service Registry, i.e. a catalogue that specified which
MEC services can be used by MEC apps, and where they are located (i.e., at the MEC host itself, or at a different one).
Examples of standardized MEC services are the Radio Network Information Service (RNIS) and the Location Service.
Every MEC service advertises its presence to all the Service Registries in a MEC system via the MEC platform manager.
Figure 3 - Simu5G as emulator
6
When a MEC app queries a Service Registry, the latter returns the (local or remote) endpoints at which the available MEC
service should be contacted. This allows MEC apps to consume MEC services located in different MEC hosts.
MEC apps can use MEC services via a standard MEC APIs, built upon RESTful APIs. REST follows a request-
response paradigm. While in the MEC environment a MEC app often acts as a client, making requests, it is sometimes
useful to allow the MEC service to notify MEC apps of some events at specific times, in a subscribe/notify approach, still
implemented based on the RESTful pattern [33]. With this approach a MEC app registers its interest in a particular
resource (subscription phase), and then, when an event occurs concerning that resource, the MEC service sends a
notification to all the subscribers (notification phase).
The MEC system level maintains a global view of the MEC hosts in a MEC system, arbitrates MEC-host resources,
and manages the lifecycle of the MEC apps, i.e. instantiation, relocation and termination. Its core element is the MEC
orchestrator. The orchestrator receives the requests for instantiation or termination of MEC apps issued by the user’s
device application, after a granting operation made by the Operation Support System of the network operator, which is
usually managed by the network operator and deals with authentications and authorizations. It then selects an appropriate
MEC host for that MEC app, based on required constraints (e.g., latency), available resources (e.g., memory, disk, CPU),
and available MEC services. Then it instructs the appropriate MEC platform manager to deploy the virtual machine that
will run the MEC app. The User application lifecycle management proxy (UALCMP) acts as a bridge between the device
application and a MEC system: the requests coming from the former are forwarded to a MEC orchestrator, and responses
from the MEC orchestrator are sent to the device application. A MEC deployment can include several MEC systems, i.e.,
orchestrators managing disjoint sets of MEC hosts.
A MEC app is onboarded through an application package. The latter is composed of a bundle of files provided by the
application provider, onboarded into the MEC system and used by the latter to instantiate an application. It also includes
the Application Descriptor describing the application requirements and rules required by the MEC app [35].
A UE interacts with the MEC system via two logically distinct entities: the device app and the UE app. The device app
interfaces with the MEC system to request specific functions related to life-cycle management of a MEC app (notably,
instantiation or termination). The UE app is instead the endpoint of data-plane communication between the UE and the
Figure 4 - Main entities in an ETSI-MEC infrastructure (taken from [34]). Black/grey modules are/are not modelled in our framework
7
Figure 5 - Sequence diagram of the instantiation of a MEC app on behalf of the user.
MEC app, which starts after the device app receives confirmation of the instantiation of the MEC app. The sequence
diagram of a MEC app instantiation is shown in Figure 5. We remark that neither the interface between the UE app and the
device app, nor the data-plane communication between the UE app and the MEC app, are part of the ETSI MEC standard.
The ETSI standard also defines a set of reference points [34], i.e. standardized interfaces between the entities of a MEC
system. For instance, the Mx2 reference point is used by the device application to request the MEC system to perform
operations on a MEC app. The UALCMP is in fact the point of termination of the Mx2 interface towards the device
application. The Mp1, between the MEC platform and the MEC app, allows MEC services management (registration and
discovery).
3. An Architecture for Rapid Prototyping of MEC-based Applications
This section presents our framework for rapid prototyping of MEC apps. Our framework is based on Simu5G, which is
enhanced with models of the MEC components described in the previous section. The rationale behind this choice (instead
of, say, a standalone model of the MEC system) is twofold: on one hand, Simu5G simulates the 4G/5G data plane, which
allows a MEC app developer to obtain a reliable estimate of the performance of her app in a mobile network. On the other
hand, Simu5G produces those data that a MEC platform uses to provide MEC services (notably, location and radio
information), hence allows a developer to test MEC apps that consume those services. A design choice for our framework
is to implement in a standard-compliant way (large subsets of) the MEC reference points towards applications, i.e. the Mx2
and Mp1 interfaces. This design choice allows a developer to interact with our framework in the same way it would with a
real MEC system: this way, one can interface production-level code with it. Since Simu5G can run as both a simulator and
an emulator, and can interface with external code, a developer has the option of writing parts of her MEC app either as
Simu5G modules (e.g., a stub UE app for testing the MEC app) or as standalone external applications. In this section, we
describe how we modelled MEC entities. In particular, we focus on our model of the MEC system-level entities, the MEC
host, MEC platform and services, and application endpoints.
3.1. Model of MEC system-level components
Our framework models both the UALCMP and the MEC orchestrator. The UALCMP is an OMNeT++ module, which
includes the TCP/IP stack and the application that implements the Mx2 reference point [36]. With this interface, device
applications can query the instantiation and the termination of MEC apps, exposed through a RESTful API. Such requests
are then forwarded to the orchestrator which, after handling them, returns a response on the outcome. The UALCMP can
8
also be queried by a device application running outside the simulator, if connected to an INET external interface. Note that
coherently with the ETSI standard we also allow a device application to join an already instantiated MEC app, in a
many-to-one communication model. In this case, the UALCMP simply retrieves and returns the endpoint of such MEC
app. For ease of exposition, we do not discuss this possibility further, assuming one-to-one communication henceforth.
The MEC orchestrator is a simple module, connected to the UALCMP. It is configured with a list of the attached MEC
hosts. Upon receiving a request for the instantiation of a MEC app, the orchestrator chooses the most suitable MEC host,
among those associated to it, according to a user-definable policy which may take into account application requirements
(e.g., CPU, memory, disk, required MEC services). Next, it contacts the MEC platform manager of the chosen MEC host
in order to eventually trigger the MEC app instantiation, following admission control. A user interested in defining or
testing MEC host selection policies can do so by overriding the chooseBestMECHost() method of the MEC
orchestrator module. The time needed to accomplish the instantiation and termination operations can also be configured.
Interactions between the MEC orchestrator and the UALCMP or the MEC host-level entities are not implemented to be
standard-compliant. Instead, they occur using OMNeT++ message passing between modules. This allows us to implement
standard-compliant behavior with a simpler, more manageable implementation, without any loss of functionality towards
the application endpoints. Moreover, we remark that not all reference points have been standardized at the time of writing,
e.g., the interface between the UALCM proxy and the Operation Support System. For this reason, we did not implement
the latter, and modelled the above interactions with a configurable delay in the MEC orchestrator.
Finally, note that our framework allows one to instantiate several MEC system, each one with its own MEC
orchestrator and set of hosts.
3.2. Model of the MEC host-level components
The MEC host is the main building block of the MEC host level architecture. As shown in Figure 6, it includes both
management entities, such as the MEC platform Manager and the Virtualisation Infrastructure Manager, and the modules
required to run the MEC apps, i.e. the virtualisation infrastructure and the MEC platform. Hereafter, we first describe how
the MEC host runs MEC apps. We defer describing the MEC platform and MEC services to the next subsection.
A MEC host comes with a set of hardware resources, namely a processing rate (measured in instructions per second), a
given amount of memory and disk space. The main duty of a MEC host is to run MEC apps. As outlined before, a MEC
app can be either external to our framework, i.e. an external application exchanging information with our framework in
real time, or internal to our framework, i.e., an OMNeT++ module compiled within it. In the external case, the MEC host
has an external interface towards the MEC app, which runs somewhere else (e.g., on a virtual machine in a desktop
computer connected to our framework via a local network). Accordingly, the speed at which that MEC app runs depends
on the hosting hardware (e.g., its CPU speed). In the internal case, i.e., when MEC apps run within our framework,
instead, we need to manage the pace a MEC app is executed. In a discrete-event simulator, such as Simu5G, simulated
time passes between scheduled events. Messages and packets are events, hence a packet exchange (e.g., a request/response
pattern) occurs over a span of simulated time. On the other hand, the MEC application logic is embedded within event
handlers, and event handling per se does not consume simulated time: any processing done by such applications is thus
instantaneous in simulated time. So, to obtain a more realistic model of the processing operations in a MEC app, we
modelled the execution time of a block of code by adding an event representing the required CPU computation, which in
turn depends on the level of CPU contention at the MEC host. A MEC host can schedule internal MEC apps according to
two paradigms:
9
Segregation, whereby each MEC app obtains exactly the amount of computing resources it has stipulated at the
time of admission control, even when no other MEC apps are running concurrently;
Fair sharing, whereby active MEC apps share all the available computing resources proportionally to their
requested rate, possibly obtaining more capacity than stipulated when contention is low.
At the time of instantiation, a device application requests a computation rate for the MEC app. The admission
control done by the orchestrator checks (among other things) that
  , where is the MEC host’s processing rate.
With segregation, a MEC app computation will be served at a rate , regardless of the processing contention on the MEC
host. This models non-work-conserving scheduling of the MEC app virtual machines on the MEC host. With fair sharing,
instead, a MEC app computation will run at a rate
 

where the summation at the denominator includes all the MEC apps that are active when the computation is requested.
It is obviously  , equality holding only when the admission control limit is reached. Fair sharing approximates
Generalized Processor Sharing (GPS) scheduling on the MEC host [30]. The approximation is due to the fact that the
actual processing rate is computed once when a computation is requested by the MEC app, rather than updated
continuously over time. In fact, changes over time (because other MEC apps start and terminate, modifying ), hence
the time at which a computation terminates cannot be calculated once and for all ex ante, but must instead be updated
every time the summation at the denominator changes. This is a well-known problem with GPS emulation, where a
quadratic complexity is required for exact tracking of finishing times. However, exact GPS emulation would only bring a
significant accuracy improvement with very long continuous computations done by a MEC app. Note that one can
instantiate a dummy MEC app on a MEC host, whose only purpose is to eat up resources (notably, processing speed) at the
MEC host, so as to simulate a given processing load in an efficient way.
To model the time spent by computation requests in an internal MEC app, we allow a user to use a special call
compute(N), where is the number of instructions to be executed. This adds an event in the future, whose firing time is
computed based on , on the MEC host processing speed, on the scheduling paradigm and possibly on the number of
10
active MEC apps at the time of the call. After such delay, the block of C++ code modelling the computation phase of an
internal MEC app is executed.
3.3. Model of the MEC platform
The MEC platform module has its own TCP/IP stack and is connected to the Virtualisation Infrastructure via an
Ethernet connection. It contains MEC services and the Service Registry, which are implemented as TCP applications.
MEC Services interface with MEC apps through a RESTful approach. In our framework, they are implemented as
HTTP servers and support both the request-response and subscribe-notify communication models. Incoming requests are
queued up in FIFO order and served sequentially. A notification FIFO queue is also added (different events may generate
notifications simultaneously, hence queueing may actually occur). When both are non-empty, the notification queue takes
precedence. A MEC service is itself a point of contention, since several MEC apps may request the same service near-
simultaneously. So, we need to model delays at a MEC service in a coherent and scalable way. The service time of an
HTTP request or a subscription notification event is simulated with a delay computed by a specific method,
calculateRequestServiceTime() that the user can modify in order to produce such times according to some
attributes, such as request type, the number of parameters, a random distribution. By default, the service time is calculated
according to a Poisson distribution with a configurable mean. Thus, the response time of an HTTP request depends on the
calculated service time and on the number of requests already queued at the server.
A user may want to test scenarios with heavy contention at a MEC service (e.g., due to a very large number of
concurrent MEC apps). In a MEC system, in fact, requests can arrive not only from the MEC apps instantiated in the same
MEC host where the MEC service runs, but also from MEC apps deployed in other MEC hosts [34]. Simulating many
MEC apps increases the computation overhead in Simu5G, and this may constrain the size of the scenario that can be
simulated or emulated in real time. To solve this, we provide an implementation which allows one to simulate arbitrary
loads at a MEC service at a constant computation cost. We distinguish foreground MEC apps, i.e., those that are
instantiated on a MEC system, and background MEC apps, i.e. those whose load we want to simulate, without paying the
overhead of modelling the application logic itself. We model a MEC service as an M/M/1 queueing system, where the
service rate is and the arrival rate of foreground/background requests are and , respectively, with  .
Moreover, we assume that foreground and background requests are independent. If   , then the system is stable,
and the state probabilities seen by an arriving foreground request are the same as the steady-state probabilities seen by a
random observer (PASTA property), which are:
   
(1)
where   , and    is the number of requests in the queue. When a foreground request arrives at the
queue, two cases are given: a) no other foreground requests are in the system, or b) there is already one foreground request
in the system. In the first case, then one can extract a value for the number of requests already in the system from (1), let it
be , and schedule the departure of the arriving foreground request   service times in the future. This entails
extracting a value from an  -stage Erlang distribution with a rate .
Assume instead that a foreground request arrives when there is already another in the system. Let   be the arriving
times of the former and current foreground requests, respectively. The number of background requests arrived in   is
a Poisson random variable with a mean   . Once a value is extracted from that distribution, call it , then one
11
can follow the same reasoning as above, and extract the inter-departure time between the two foreground requests from an
 -stage Erlang distribution with a rate .
By using the above model, one can store only foreground requests in the queue, hence the computation overhead of
simulating contention at the MEC service is independent of the background load. If the arrivals and/or services are not
exponential, a similar modelling can be used, by substituting the appropriate distributions for the number of requests in the
system, the cumulative service time of   requests back-to-back, and the number of arrivals within a known interval of
time.
To validate the above model, we simulated a simple scenario with a single gNB, a MEC host and three mobile UEs
having MEC apps making periodic requests to the Location Service running in the MEC host. The Location Service is
loaded by other MEC apps, simulated either as real MEC apps modules or using the above load generator. These send
requests exponentially with a rate lambda = 0.024 each, while the three foreground apps send requests every 500ms. The
number of background MEC apps varies from 10 to 300 and the simulation time is 180 seconds.
Figure 7 shows the CDF of the response time of foreground requests issued by the foreground MEC apps when the
number of background MEC apps varies. As we can see, results are overlapping even though requests from foreground
MEC apps are periodic, instead of Poisson. However, the computation overhead is quite different. We show in Figure 8 the
mean execution time out of 15 independent repetitions, with 95% confidence interval, in the same conditions, on a laptop
with an Intel I5 processor. With the load generator, the running time is independent of the number of background MEC
apps, and remains fixed at 3.8 seconds. When simulating real MEC apps modules, the running time increases with the
number of background MEC apps. Note that, already with 200 real MEC apps, the mean wall-clock time required to run a
simulation of 180 seconds exceeds 200 seconds, i.e., simulation time runs slower than wall-clock time. This means that
emulation would not be workable in that case. By using the above load generator, instead, one could emulate in real time
scenarios where MEC services are congested.
4. MEC Applications and Services
In this section, we describe how we model the MEC app, device app and UE app application endpoints, showing what
configurations are required depending on the possible deployment options (i.e., internal to Simu5G or real applications
running outside Simu5G). Moreover, MEC services running within the MEC platform are described, presenting the
implementation of two ETSI MEC services.
Figure 7 - CDF of the delay of the foreground MEC apps when an
increasing number of real/modelled MEC apps interfere at the MEC
service.
Figure 8 - Wall-clock time required to simulate 180s of a MEC system,
using an increasing number of MEC applications vs. the equivalent load
generator
12
4.1. Model of application endpoints
As already mentioned, Simu5G can also exchange real network packets with external applications. Since our MEC
framework implements the Mx2 and Mp1 reference points, a developer can develop and test all of the MEC-related
applications as either Simu5G modules, or external applications. In this last case, she can leverage the fact that Simu5G
can run in real time to setup live experiments in an emulated environment.
4.1.1. MEC app
Each MEC app must be accompanied by the JSON file describing the Application Descriptor mentioned in Section 2.
This file includes the necessary fields to allow on-boarding of an application package into the MEC system, such as:
appDId: unique identifier of the app descriptor;
appName: name of the MEC app;
appProvider: module name necessary for create the OMNeT++ module, if the MEC app is internal the simulator, or it
can be left empty if the MEC app is external;
appServiceRequired: the MEC services needed to run the MEC app;
virtualComputeDescriptor: the computation resources required to run the MEC app in a MEC host (i.e. memory, disk
and CPU).
If the MEC app is external to Simu5G, a field named emulatedMecApplication must be specified. This contains the IP
address and port sub-fields identifying the real MEC app endpoint. This way, the MEC orchestrator is made aware that the
MEC app to be instantiated runs outside Simu5G, hence it returns the MEC app’s endpoint to the device app, instead of
instantiating it inside the simulator.
Our framework allows a developer to quickly create a prototype of an internal MEC app, by deriving the MecAppBase
module. The latter is a base class that manages sockets and OMNeT++ events, leaving to the developer only the
implementation of the methods called upon message reception (e.g. from the Service Registry, a MEC service or a UE).
For external MEC apps, on the other hand, the only required configuration is the Service Registry’s IP address and port,
through which the location (i.e. IP and port) of the required MEC services can be discovered.
4.1.2. Device app
To facilitate the task of an application developer, our framework provides a simple device app that can request the
installation and termination of a MEC app to the UALCM proxy via the RESTful API implementing the Mx2 reference
point. That device app can be contacted by a UE app via UDP socket by means of a simple interface that includes
messages for the creation, termination and acknowledgments of a MEC app (e.g. START mecAppName, ACK
endpoint). This interface can be used by internal and external UE apps, thus a developer has only to write the UE app
logic and simply query the above device app when a MEC app instantiation (or termination) is needed.
Note that our UALCM proxy also accepts requests from external device apps, since the interface between the two
occurs via the ETSI compliant Mx2 API. In this case, the external device app only needs to be configured with the address
of the UALCM proxy.
4.1.3. UE app
As far as the UE app is concerned, the same approach as for the other apps has been used, i.e. it can be both internal
and external to Simu5G. The first option can be particularly useful when the UE app is meant to run as a stub, e.g. to just
issue requests at a predefined rate and record statistics.
13
Figure 9 describes the UE app pseudocode necessary to request the instantiation (and termination) of the MEC app, via
the built-in device app available in Simu5G. The above design makes it very easy for a developer to port to a MEC
environment a preexisting client-server application: the existing server application can be deployed via the MEC
framework as a MEC app, whereas the entire client logic is inserted virtually as is within the curly brackets in Figure 9
with minimal to null modifications.
4.2. Model of MEC services
As far as MEC services are concerned, we provide both a general-purpose module for rapidly prototyping ETSI-
compliant MEC services, and two useful standard services, namely RNIS and Location Service. Our framework comes
with a basic module, called MecServiceBase, which implements all the non-functional requirements needed for
running an HTTP server, leaving to a developer only to implement the methods to handle HTTP requests and subscriptions
of the RESTful API. The interface also maintains the set of eNB/gNB modules associated to the MEC host.
Two standardized MEC services are currently implemented, namely the RNIS and the Location Service. The RNIS is
used to gather up-to-date information regarding radio network condition and the UEs connected to the base station
associated to the MEC hosts [31]. Such information allows MEC apps to have real-time information on the network
performance, which can be used e.g., to provide an improved Quality of Experience (QoE) to end-users: for instance, to
decrease the video quality of a streaming application while the channel is poor or the cell is overloaded, thus avoiding
video rebuffering. A subset of the RNIS API is implemented, which includes the Layer-2 measures resource, reporting
gNB Layer 2 measurements, such as packet delay, throughput, number of active UEs with downlink/uplink traffic, data
volumes, cell utilization or packet data rate. By varying the network configuration, one can carry out evaluations and
validations of MEC apps using the RNIS in different network conditions. The RNIS gets its information from gNB
modules within Simu5G. Hereafter we show how this is achieved without modifying the existing models within Simu5G.
A dedicated module, named gNodeBStatsCollector (collector, from now on), can be added to a gNB to retrieve
measures from the NIC modules. This should be instantiated only when necessary (i.e., when a MEC host exposes the
RNIS), as it is costly from a processing overhead standpoint. The large overhead is due to the fact that a collector manages
several timers used to trigger information-retrieving procedures, and their firing events slow down the execution time.
Each logged measure is stored in a L2Measure object by the collector. Different aggregators can be used to compute the
value to be returned when the RNIS requires it, e.g. average, moving average, or last sample. A user can also implement its
own aggregator or configure window timers according to documents [37]-[38]. Some L2 measures of the RNIS involve
different sublayers of the NR stack. For instance, some delays are measured from the arrival to the PDCP layer to the
Figure 9 - Pseudo-code to execute a UE app in a MEC system
14
reception of the HARQ ACK at the MAC layer. This is handled without modifications to the layers, by adding a new
module in the NIC, called packetFlowManager, which receives information by the relevant layers when the above
events are triggered and maintains the data structures necessary to identify the same payload at different NR sublayers.
We validate the results returned by our implementation of the RNIS by comparing them against similar metrics already
provided by Simu5G. Figure 10, top, shows the PDCP packet delay in downlink, returned by the
packetFlowManager, of a UE that connects to three gNBs over time, the middle one having other 100 UEs
periodically receiving packets from a server. In order to show readable and interpretable results, the number of available
5G resource blocks has been downsized to 10. When the UE handovers to the middle gNB (handovers are marked by
dashed vertical lines), the delay increases. Figure 10, center, depicts instead the PDCP delay calculated by Simu5G in the
same simulation. It can be observed that the patterns are correlated, suggesting a correct implementation of the RNIS
metric. A closer look, shown in Figure 10, bottom, and reporting a zoom of the two above graphs in a single reference,
shows that there is a constant difference of 1ms, which is due to the fact that the RNIS uses HARQ ACK reception to mark
the end of the measurement interval, and that occurs 1ms after the UE has received the packet over the air, the latter being
the event when the Simu5G statistic is computed instead. Moreover, delay at handover times are different (as per the
spikes in Figure 10, center, around   ,   ), since Simu5G starts measuring the delay at the entry within the NR
Figure 10 - Downlink packet delay obtained by querying the RNIS (top) or using Simu5G metrics (center), and zoom-in of the difference between the
two (bottom) between 12 and 15 seconds.
15
stack, whereas the RNIS starts at the entry at the PDCP layer. The two are usually the same, except during handover,
when Simu5G adds the handover delay to the computation.
The Location Service provides accurate information about UE and/or base station position, enabling active device
location tracking and location-based service recommendation. The reference API is described in [32] and it is based on the
RESTful API originally defined by the Open Mobile Alliance [39]. UE positions are stored and periodically updated on
the gNBs to which they are currently connected and are expressed as three-dimensional Euclidian coordinates provided by
the INET Mobility model library. A MEC app can query the position of the UEs with different granularities: a single UE, a
group of UEs, only the UEs connected to a specific base station or a group of them. The “UE Area subscription” is also
present [32]. It follows the “Area (circle) location notification" subscription defined in [39], on which a MEC app
subscribes to receive notifications when a UE enters or leaves a circular zone described by its center coordinates and
radius.
5. A Case Study
We describe here an example of a distributed application, where both endpoints (UE app and MEC app) are external.
In our scenario, a vehicular UE moves in a simulated floorplan, and wants to be notified when it enters a danger zone, i.e.
a black-ice area. To do so, it sets up a MEC app that uses the Location Service to check the UE position. The network
scenario is composed of a gNB with some associated UEs, i.e. cars equipped with a NR interface, moving towards the
danger zone. A MEC host is attached to the gNB and runs the Location Service. The latter provides the "Area (circle)
notification subscription” resource [32]. On that resource, a client requests to monitor when a UE enters in a circular zone.
More specifically, after the MEC app instantiation, the UE app requests it to monitor a specific zone. We assume the
UE starts outside the danger zone, hence the first subscription is for the entering-danger zone event. When the MEC app
receives the entering notification, it modifies its subscription, to be notified when it leaves the danger zone. The UE app is
only informed upon notification events. The resulting sequence diagram is shown in Figure 11.
Figure 11 - Sequence diagram of the danger zone Warning Alert use
16
We run the above scenario in an emulated environment on which both the UE app and the MEC app run outside the
simulator, on the same host where Simu5G runs, as shown in Figure 12. The device app, instead, resides on the UE module
inside Simu5G. We now describe the configurations needed to execute this testbed on a computer running Linux Ubuntu
operating system.
First of all, the network scenario must be configured within Simu5G. For the low-level details, we refer the interested
reader to papers describing Simu5G configuration (e.g., [6] and [1] describing configuration of MEC scenarios) as well
as the website documentation [7]. Here, we limit ourselves to recalling that Simu5G modules can exchange packets with
outside applications using ExtLowerEthernetInterface OMNeT++ modules. These must be inserted in the car and mecHost
modules, so as to inject/receive UE app and MEC app traffic inside the simulator. In our case, these interfaces are created
as Virtual Ethernets (veth).
Then, the host operating system must be instructed to route packets into Simu5G via the veth interfaces by adding ad-
hoc routing rules to reach the device app and the MEC platform, and to enable the transport of the traffic between the UE
app and the MEC app through the emulated 5G network. Since the MEC app and the UE app run on the same host, a
mechanism to bypass the operating system and steer the traffic towards the simulator is needed. A possible solution is to
insert a natRouter module in the Simu5G network. This way, both real applications send packet to the IP addresses of the
interfaces of the natRouter, which in turn performs Network Address Translation by changing the destination addresses to
the proper real application’s addresses. Figure 13 shows the commands required to create, configure and add routes of the
veth interfaces on a host equipped with Linux Ubuntu 18.04 operating system. Moreover, routing information within the
Figure 12 - Simu5G set up for emulation
Figure 13 - veth interfaces configuration
17
simulator have to be added, as the OMNeT++ platform does not know autonomously how to route packets to destinations
outside the simulated/emulated network.
After the networking has been set, the last part to be configured is related to the MEC app in use, which is the
MECWarningAlertApp in this case. An appDescriptor file describes the MEC app in terms of its configuration parameters.
For example, for the purpose of the emulation, the emulatedMecApplication field informs the MEC orchestrator about the
IP address and port where the MEC app listens for UE app requests. Once the UE app and the MEC app know the device
app and the Service Registry endpoints respectively, the testbed is ready to run.
Finally, Figure 14 shows the timeline of the car moving towards the danger zone, with the relative events captured by
the Wireshark tool monitoring the veth0 and veth2 interfaces.
As can be seen, the configuration effort is modest, and mostly is focused on the networking settings. This is necessary
for any scenario where OMNeT++ is run in emulated mode and does not depend on the MEC framework. The latter only
requires the creation of the descriptor file for the MEC app and the creation of the network scenario to be executed.
6. Related Works
In this section we review the available works on 5G simulation, MEC simulation and MEC services testing.
As far as 5G simulators are concerned, there are both those that simulate the physical layer, such as [14]-[15], and
those that allow end-to-end, application-level simulations, like Simu5G. In this last category we find 5G-Lena [16] and
5G-air-simulator [17]. As far as we know, neither supports MEC or allow real-time emulation of a 5G network.
A critical review of the Fog/Edge simulators and emulators available to date is reported in [24]. The tools described in
the paper have been developed to evaluate computing infrastructures in terms of deployment scenarios, energy
consumption and operational costs. They address Internet of Things application and resource management in an
Edge/Fog/Cloud continuum. The most widely used seems to be iFogSim, although it does not take network aspects into
account too much. The underlying network condition, with increasing QoS requirements from users, has become of
paramount importance and therefore needs to be carefully considered as well. More importantly, these tools do not address
ETSI MEC and the related functionalities, such as MEC services, that can retrieve information about the underlying
network for the edge applications, i.e. MEC apps.
Figure 14 - Timeline of the message events during the execution of the use case
18
The work most closely related to ours seems to be the MEC simulator described in [23]. Devised for the ns3
framework, it allows one to simulate simplified models of distributed MEC-based applications over a MEC system, using
non ETSI-compliant interfaces. The simulator includes models of MEC orchestrator and app mobility. There are, however,
several crucial differences with our work: first, one cannot use real MEC-based applications and run them on this
framework, as it does on Simu5G. Second, there seems to be no models for MEC services and MEC platform: the
simulator, in fact, includes dummy base stations, whose only role is to have UEs associated to them on a proximity basis,
without modelling radio communication (in fact, RAN latency in the experiments described in [23] is modelled as a
probability distribution). Therefore, on one hand, these base stations cannot provide any radio information to a MEC
platform (e.g., the level of resource occupation, or the channel quality of a mobile); on the other hand, such an architecture
also prevents straightforward interoperability with the 4G/5G libraries already available for ns3 (such as 5G-Lena), where
the above information could be generated. Last, but not least, to the best of our knowledge, [23] cannot run real-time
emulations.
All the above works focus on MEC modelling, and either neglect or abstract away the role of the underlying RAN. We
remark that the interplay between the RAN and the MEC does not limit to the former transporting packets generated or
consumed by the latter. Rather, and no less importantly, the RAN generates the information that MEC apps use via MEC
services (notably, those returned by the RNIS). While it could make at least some sense to abstract away the modelling of
RAN communication impairments (e.g., just a delay distribution and a bandwidth), it is just pointless to have the RNIS
without an underlying credible and detailed model of the RAN. We also remark that an architecture including both the 5G
access and the MEC system allows one to test the impact of MEC traffic on the network.
The ETSI group also provides a portal for exploring the functionalities of MEC services. It is called “ETSI MEC
Sandbox” [41] and offers an environment on which the users can choose among different real-time access network
scenarios (e.g. 4G, 5G, WiFi) and practice with ETSI MEC service APIs. RESTful resources can be queried from both
browser and existing applications testing MEC application use cases. However, the available settings are limited and the
network's parameters, like fading, cell interferences and multi-carrier components cannot be modified in order to produce
different behaviors, limiting the use of this tool to the sole purpose of experimenting with the MEC services API, without
allowing one to customize the network scenarios.
The OpenAirInterface 5G-RAN project [22] is under way as we write, with a schedule foreseen to complete in the
second half of 2022. The project promises to deliver an open-source software implementation of a programmable 5G
RAN. If and when this project is completed, it will probably be a good tool to experiment with the interplay between MEC
and the RAN: for instance, one may envisage building a system where OAI 5G-RAN interacts with a MEC infrastructure
(e.g., Intel OpenNESS [28]), and UEs can enjoy MEC services in an emulated environment. However, this requires
OpenAirInterface 5G-RAN to interface with the MEC system to export the information needed for MEC services, and we
have no information that this is included in the current plan of the OpenAirInterface consortium. Even if it was, we are
somewhat skeptical that such an environment would provide a MEC app developer with the possibility to test its own
MEC apps quickly and under controlled conditions, possibly including large-scale communication or congestion at the
MEC host.
Finally, three previous works of ours [13],[2],[27] are related to this one. Work [13] describes the MEC model
developed for SimuLTE. This work is meant to be used for research purposes, i.e. to allow a SimuLTE user to instantiate
simplified models of MEC-based applications similarly to [23], and evaluate their impact on the data plane. While the
19
work described in this paper reuses code from [13], suitably adapted to the 5G environment, its purpose is to serve MEC
developers, rather than networking researchers. Accordingly, it has been re-engineered and enhanced with models of
ETSI-MEC components, such as the UALCMP or the device app, that are necessary for seamless emulation of MEC
functionalities for a developer, and standard-compliant external interfaces to the application world have been implemented.
For instance, the use case described in Section 5 could not have been emulated in that environment not even substituting
4G access for 5G. Our paper [2] describes the real-time emulation capabilities of an early release of Simu5G. As a case
study, we show therein that Simu5G can provide 5G transport to a MEC app running on Intel OpenNESS [28]. On one
hand, the current version of Simu5G is quite different from the one described in [2], especially for what concerns real-time
emulation (thanks to the new versions of OMNeT++ and INET). On the other hand, the MEC app described therein is a
simple client/server video application running on a virtual machine instantiated on a MEC host. All the management-plane
interactions are missing, and the service is statically configured offline. In [27], we used Simu5G together with and Intel®
CoFluent™ studio [26] to evaluate the performance of MEC apps running over different 4G/5G deployments. CoFluent is
a modelling and simulation tool for optimizing, analyzing and predicting the performance of complex systems, that models
and simulates HW/SW systems with microinstruction-level accuracy. In [27], the MEC host was modelled within
CoFluent, and a video-streaming MEC app was modelled within it. However, the co-simulation framework described in
[27] is hardly comparable to what we describe in this paper. In fact, it is based on file exchanges between Simu5G and
CoFluent, which run separately. This allows one to compute the round-trip delay, including both communication and
computation. However, no feedback between the two simulators can exist (e.g., network conditions influencing the MEC
app behavior, e.g. via RNIS, or user behavior depending on MEC app results). This could only be enabled by scheduling
both CoFluent and Simu5G events in the same unified framework, which would require a considerable amount of work.
Moreover, while CoFluent can model a MEC host with high accuracy, its very accuracy makes it unsuitable to model a
large-scale MEC system, which can instead be modelled with Simu5G.
7. Conclusions and Future Work
This work described a framework for rapid prototyping of MEC-based applications. Our framework, based on the
Simu5G discrete-event simulator, gives a developer two options: the first one is to write application logic (UE app and
MEC app) as Simu5G modules. This is quite simple, and allows the developer to test the application logic in a pre-
production stage, obtaining reliable performance metrics in a customizable 5G scenario. The other option is to use existing
MEC-based application endpoints, and run them through our framework, which provides not only 5G packet transport, but
also MEC signaling functionalities and MEC services. This can also be done in real time, e.g. for demonstration purposes
or to test interactions with a human end user or other external software. We have described the modelling of the MEC
components in our framework, validated our implementation of MEC services, and showed that one can set up an
emulation testbed with external application quite simply, on off-the-shelf hardware. We believe that the work described in
this paper will be useful to MEC app developers. To the best of our knowledge, there are no tools with similar
functionalities available to the developer community.
At the time of writing, the above framework is being used in the framework of the Hexa-X EU project [29]. In that
framework, it will support the development, validation, evaluation and demonstration of federated learning of explainable
AI models. More specifically, we plan to use our framework to evaluate network protocols for federated learning, where
learning logic can run on both UEs and in MEC systems.
20
Acknowledgments
This work was partially supported by the Italian Ministry of Education and Research (MIUR) in the framework of the
Cross-Lab project (Departments of Excellence), and by the European Commission through the H2020 projects Hexa-X
(Grant Agreement no. 101015956).
References
[1] A. Noferi, G. Nardini, G. Stea, and A. Virdis, “Deployment and configuration of MEC apps with Simu5G”, 8th OMNeT++ Community Summit
2021, virtual, September 8-10, 2021, ArXiv abs/2109.12048
[2] G. Nardini, G. Stea, A. Virdis, D. Sabella, P. Thakkar, "Using Simu5G as a Realtime Network Emulator to Test MEC Apps in an End-To-End 5G
Testbed", PiMRC 2020, London, UK, 1-3 September 2020
[3] G. Nardini, G. Stea, A. Virdis, “Scalable Real-time Emulation of 5G Networks with Simu5G”, IEEE Access, 2021, DOI:
10.1109/ACCESS.2021.3123873
[4] M. Carson, D. Santay, "NIST Net: a Linux-based network emulation tool", SIGCOMM Computer Communication Review 33, 3 (July 2003),
111126. DOI: https://doi.org/10.1145/956993.957007
[5] P. Imputato, S. Avallone, "Enhancing the fidelity of network emulation through direct access to device buffers", Journal of Network and
Computer Applications, vol. 130, 2019, pp. 63-75, ISSN 1084-8045, DOI: https://doi.org/10.1016/j.jnca.2019.01.014.
[6] G. Nardini, D. Sabella, G. Stea, P. Thakkar, A. Virdis "Simu5G An OMNeT++ library for end-to-end performance evaluation of 5G networks",
IEEE Access, vol. 8, pp. 181176-181191, 2020, DOI: 10.1109/ACCESS.2020.3028550
[7] Simu5G Website, http://simu5g.org, accessed October 2021.
[8] OMNeT++ Website: https://omnetpp.org, accessed October 2021.
[9] INET Library Website. https://inet.omnetpp.org, accessed October 2021.
[10] M. Hu, X. Luo, J. Chen, Y.C. Lee, Y. Zhou, D. Wu, "Virtual reality: A survey of enabling technologies and its applications in IoT", Journal of
Network and Computer Applications, vol. 178, 2021, ISSN 1084-8045, DOI: https://doi.org/10.1016/j.jnca.2020.102970.
[11] C. Sommer, R. German, F. Dressler, "Bidirectionally Coupled Network and Road Traffic Simulation for Improved IVC Analysis," IEEE
Transactions on Mobile Computing (TMC), vol. 10 (1), pp. 3-15, January 2011
[12] 3GPP TR 38.801 v14.0.0, “Study on new radio access technology: Radio access architecture and interfaces (Release 14)”. March 2017
[13] G. Nardini, A. Virdis, G. Stea, A. Buono, "SimuLTE-MEC: extending SimuLTE for Multi-access Edge Computing", 5th OMNeT++ Community
Summit 2018, Pisa, Italy, 5-7 September 2018
[14] Y. Kim et al., "5G K-Simulator: 5G System Simulator for Performance Evaluation," 2018 IEEE International Symposium on Dynamic Spectrum
Access Networks (DySPAN), Seoul, 2018, pp. 1-2, doi: 10.1109/DySPAN.2018.8610404.
[15] M. Müller, F. Ademaj, T. Dittrich, et al. "Flexible multi-node simulation of cellular mobile communications: the Vienna 5G System Level
Simulator". J Wireless Comm. Network 2018, 227 (2018). doi.org/10.1186/s13638-018-1238-7
[16] N. Patriciello, S. Lagen, B. Bojovic, L. Giupponi, "An E2E simulator for 5G NR networks", Simulation Modelling Practice and Theory, Volume
96, 2019, doi.org/10.1016/j.simpat.2019.101933.
[17] S. Martiradonna, A. Grassi, G. Piro, and G. Boggia,"5G-air-simulator: an open-source tool modeling the 5G air interface", Computer Networks
(Elsevier), 2020.
[18] Simnovus Callbox website: https://simnovus.com/products/callbox/, accessed October 2021.
[19] Viavi solutions website, available at: https://www.viavisolutions.com/, accessed October 2021.
[20] Polaris Networks 5G emulators website: https://www.polarisnetworks.net/5g-network-emulators.html, accessed October 2021.
[21] Keysight technologies website, available at https://www.keysight.com/, accessed October 2021.
[22] OpenAirInterface 5G Ran project website https://openairinterface.org/oai-5g-ran-project/, accessed October 2021.
[23] S. Massari, N. Mirizzi, G. Piro, and G. Boggia,"An Open-Source Tool Modeling the ETSI-MEC Architecture in the Industry 4.0 Context", Proc.
of IEEE Mediterranean Conference on Control and Automation (MED), 2021
[24] A. Aral, V. De Maio, “Simulators and Emulators for Edge Computing”, in: Edge Computing: Models, technologies and applications, Editors: J.
Taheri, S. Deng, IET. 2020
21
[25] A. Virdis, G. Stea, G. Nardini, “Simulating LTE/LTE-Advanced Networks with SimuLTE”, in Obaidat M.S., Kacprzyk J., Oren T., Filipe J. (eds)
“Simulation and Modeling Methodologies, Technologies and Applications”, Springer, 2015
[26] Intel CoFluent Studio, available at: https://www.intel.com/content/www/us/en/cofluent/overview.html, accessed October 2021.
[27] A. Virdis, G. Nardini, G. Stea, D. Sabella, "End-to-end performance evaluation of MEC deployments in 5G scenarios", MDPI Journal of Sensor
and Actuator Networks, 9(4), 57, 2020, DOI: 10.3390/jsan9040057
[28] Intel OpenNESS, available at: https://www.openness.org, accessed October 2021.
[29] Hexa-X European Project: https://hexa-x.eu, accessed October 2021.
[30] A. K. Parekh and R. G. Gallager, “A generalized processor sharing approach to flow control in integrated services networks: The single node
case,” IEEE/ACM Trans. Networking, vol. 1, pp. 344–357, June 1993.
[31] ETSI GS MEC 012 v2.1.1, “Mobile Edge Computing (MEC); Radio Network Information API”, 2019
[32] ETSI GS MEC 013 v2.1.1, “Mobile Edge Computing (MEC); Location API”, 2019
[33] ETSI GS MEC 009 v3.1.1, ”Multi-access Edge Computing (MEC); General principles, patterns and common aspects of MEC Service APIs”,
2021
[34] ETSI GS MEC 003 v2.2.1, “Multi-access Edge Computing (MEC); Framework and Reference Architecture”, 2020
[35] ETSI GS MEC 010-2 v2.1.1, “Multi-access Edge Computing (MEC); MEC Management; Part 2:Application lifecycle, rules and requirements
management”, 2019
[36] ETSI GS MEC 016 v2.2.1, "Multi-access Edge Computing (MEC); Device application interface", 2020
[37] ETSI TS 136 314 v15.1.0, "LTE; Evolved Universal Terrestrial Radio Access (E-UTRA); Layer 2 Measurements (3GPP TS 36.314 version
15.1.0 Release 15) ", 2018
[38] ETSI TS 136 314 v16.0.0, "5G; NR; Layer 2 measurements (3GPP TS 38.314 version 16.0.0 Release 16)", 2020
[39] OMA-TS-REST-NetAPI-TerminalLocation-V1-0-1-20151029-A: "RESTful Network API for Terminal Location"
[40] OMA-TS-REST-NetAPI-ZonalPresence-V1-0-20160308-C: "RESTful Network API for Zonal Presence"
[41] ETSI Sandbox Website: https://try-mec.etsi.org, accessed October 2021.
[42] Location API sWebsite: https://networkbuilders.intel.com/commercial-applications/links-foundation, accessed October 2021.
ResearchGate has not been able to resolve any citations for this publication.
Conference Paper
Full-text available
Multi-access Edge Computing (MEC) is expected to act as the enabler for the integration of 5G (and future 6G) communication technologies with cloud-computing-based capabilities at the edge of the network. This will enable low-latency and context-aware applications for users of such mobile networks. In this paper we describe the implementation of a MEC model for the Simu5G simulator and illustrate how to configure the environment to evaluate MEC applications in both simulation and real-time emulation modes.
Chapter
Full-text available
In this chapter, we perform a study of the existing tools for the evaluation of Fog/Edge infrastructures. First, we analyze the state-of-the-art in the simulation of Fog/Edge infrastructures and determine the main challenges in simulation and modeling such infrastructures. Then, we use a scientific methodology to identify the most important simulation and emulation tools, identifying their main characteristics, and define a classification. Each tool is then described in detail, and compared with the others. Finally, we conclude the chapter with a discussion about future research directions in the area.
Article
Full-text available
Multi-access Edge Computing (MEC) promises to deliver localized computing power and storage. Coupled with low-latency 5G radio access, this enables the creation of high added-value services for mobile users, such as in-vehicle infotainment or remote driving. The performance of these services as well as their scalability will however depend on how MEC will be deployed in 5G systems. This paper evaluates different MEC deployment options, coherent with the respective 5G migration phases, using an accurate and comprehensive end-to-end (E2E) system simulation model (exploiting Simu5G for radio access, and Intel CoFluent for core network and MEC), taking into account user-related metrics such as response time or MEC latency. Our results show that 4G radio access is going to be a bottleneck, preventing MEC services from scaling up. On the other hand, the introduction of 5G will allow a considerable higher penetration of MEC services.
Article
Full-text available
In this paper we introduce Simu5G, a new OMNeT++-based model library to simulate 5G networks. Si-mu5G allows users to simulate the data plane of 5G New Radio deployments, in an end-to-end perspective and including all protocol layers, making it a valuable tool for researchers and practitioners interested in the performance evaluation of 5G networks and services. We discuss the modelling of the protocol layers, net-work entities and functions, and validate our abstraction of the physical layer using 3GPP-based scenarios. Moreover, we show how Simu5G can be used to evaluate Multi-access Edge Computing (MEC) and Cellu-lar Vehicle-to-everything (C-V2X) services offered through a 5G network
Article
Real-time emulation of 5G networks is highly beneficial for several purposes, such as prototyping or performance evaluation of distributed applications meant to run on 5G networks, research demonstration, evaluation of other technologies (e.g., Multi-access Edge Computing) meant to interoperate with 5G access. In this work, we describe how to use Simu5G, a new end-to-end simulator of 5G networks based on OMNeT++, as a real-time emulator. We describe in detail the modeling choices that allow emulation to scale up without compromising accuracy. We present a thorough evaluation of the Simu5G’s emulation capabilities, showing that networks with hundreds of simulated users and tens of cells can be emulated on a single desktop machine.
Article
Virtual Reality (VR) has shown great potential to revolutionize the market by providing users immersive experiences with freedom of movement. Compared to traditional video streaming, VR is with ultra high-definition and dynamically changes with users’ head and eye movements, which poses significant challenges for the realization of such potential. In this paper, we provide a detailed and systematic survey of enabling technologies of virtual reality and its applications in Internet of Things (IoT). We identify major challenges of virtual reality on system design, view prediction, computation, streaming, and quality of experience evaluation. We discuss each of them by extensively surveying and reviewing related papers in the recent years. We also introduce several use cases of VR for IoT. Last, issues and future research directions are also identified and discussed.
Conference Paper
Multi-access Edge Computing (MEC) allows users to run appli-cations on demand near their mobile access points. MEC appli-cations will exploit 5G infrastructure, and they will have to be designed by taking into account the characteristics of 5G mobile networks. This work describes how to use a system-level simula-tor of 5G networks – namely Simu5G, which evolves the popu-lar 4G network simulator SimuLTE – as a real-time 5G network emulator. This allows designers of networked applications – and MEC ones in particular – to use it as a testbed during the de-ployment. We describe the system setup of Simu5G as an emula-tor, and its emulation capabilities and scale. Moreover, we pre-sent a case study of a MEC testbed using Intel’s Open Network Edge Services Software (OpenNESS) toolkit, based on a recent demonstration in 5GAA (5G Automotive Association).
Article
5G-air-simulator is an open-source and event-driven tool modeling the key elements of the 5G air interface from a system-level perspective. It implements several network architectures with multiple cells and users, different mobility and application models, a calibrated link-to-system model for physical and data-link layers, and a wide range of features standardized for both control and user planes, as well as a set of technical components recently designed for the 5G air interface (such as massive Multiple Input Multiple Output, extended multicast and broadcast transmission schemes, predictor antennas, enhanced random access procedure, and NB-IoT). The tool has been already used in different research activities to design and evaluate the performance of reference 5G-enabled use cases. Moreover, it allows a flexible configuration, arrangement, and extension of its capabilities to model both new scenarios and new technical components, hence supporting advanced studies willing to address the research questions emerging from the deployment of current and upcoming mobile systems.
Article
Researchers from academia, industry and research centers often resort to emulation to overcome the drawbacks associated with network simulation and experimental evaluation. Emulation is broadly classified in environment emulation, usually carried out by running real code in Virtual Machines (VMs) or containers, and network emulation, typically involving network simulators that exchange packets with the real world. In this paper, we focus on network emulation, which is often exploited for rapid prototyping and testing of network protocols and algorithms. We identify the limitations of the approach currently used by various network simulators to provide network emulation and design an alternative solution based on netmap, a framework for high speed packet I/O which is available on multiple operating systems. We argue that the proposed solution to network emulation provides extremely accurate results in terms of packet latency and packet drops and prove our claim by means of an extensive experimental campaign. We also show that by building upon an accurate network emulation mechanism it is possible to validate the implementation of protocols found in network simulators against their implementation in real network stacks. As an example, we present the results of the experiments we conducted to validate the ns-3 implementation of various packet schedulers against their Linux counterpart.