Adaptive Nature-inspired Fog Architecture

Dragi Kimovski Humaira Ijaz Nishant Saurabh

Distributed multimedia systems group Distributed and parallel systems group Distributed and parallel systems group
University of Klagenfurt University of Innsbruck, Austria University of Innsbruck
Klagenfurt, Austria University of Sargodha, Pakistan Innsbruck, Austria
Email: [email protected] Email: [email protected] Email: [email protected]

Radu Prodan
Distributed multimedia systems group
University of Klagenfurt
Klagenfurt, Austria
Email: [email protected]

Abstract—During the last decade, Cloud computing has ef- In the last decade, Cloud computing has efficiently exploited
ficiently exploited the economy of scale by providing low cost the economy of scale by providing low cost computational
computational and storage resources over the Internet, eventually and storage resources over the Internet, eventually leading to
leading to consolidation of computing resources into large data
centers. However, the nascent of the highly decentralized Internet consolidation of computing resources into large data centers.
of Things (IoT) technologies that cannot effectively utilize the However, the nascent of the highly decentralized IoT tech-
centralized Cloud infrastructures pushes computing towards re- nologies that cannot effectively utilize the centralized Cloud
source dispersion. Fog computing extends the Cloud paradigm by infrastructures pushes computing towards resource dispersion.
enabling dispersion of the computational and storage resources For example, autonomous vehicles cannot rely on the Cloud
at the edge of the network in a close proximity to where
the data is generated. In its essence, Fog computing facilitates for real-time video processing or temporary data storage, as
the operation of the limited compute, storage and networking this induces unacceptably high decision making latencies.
resources physically located close to the edge devices. However, Essentially, these applications require the processing and data
the shared complexity of the Fog and the influence of the storage to be moved from the remote Cloud to the nearby
recent IoT trends moving towards deploying and interconnecting edge of network, allowing low latency communication and
extremely large sets of pervasive devices and sensors, requires
exploration of adaptive Fog architectural approaches capable processing.
of adapting and scaling in response to the unpredictable load In the digital world, the connectivity provided by the
patterns of the distributed IoT applications. In this paper we Internet network can lead us to a false sense of proximity.
introduce a promising new nature-inspired Fog architecture, For example, Cloud services can be perceived by the end-
named SmartFog, capable of providing low decision making users as being logically in a close proximity, even though the
latency and adaptive resource management. By utilizing novel
algorithms and techniques from the fields of multi-criteria physical distance can be stretched over different continents,
decision making, graph theory and machine learning we model resulting in higher end-to-end latency and lower available
the Fog as a distributed intelligent processing system, therefore bandwidth. Fog computing extends the Cloud paradigm by
emulating the function of the human brain. enabling dispersion of the computational and storage resources
at the edge of the network in a close proximity to where the
I. I NTRODUCTION data is generated [1], [2], [3]. In its essence, Fog computing
The recent embrace of the “Smart Anything Everywhere” facilitates the operation of the limited compute, storage and
paradigm has caused a major technological tidal wave in networking resources physically located close to the edge
pervasive computing, transforming the way we perceive, uti- devices. The proximity characteristics of the Fog paradigm
lize and interact with the environment around us. Large- pushes the evolution of the Cloud for the future IoT systems by
scale Internet of Things (IoT) systems, such as smart cities, enabling highly responsive services, improving the scalability
autonomous vehicles and intelligent health care services, are to new dimensions, and providing much higher fault tolerance
clearly the next disruptive technology encompassing various by masking Cloud outages [4].
physical and virtual loosely connected devices interacting There exist various overlapping definitions of Fog comput-
through existing communication infrastructure. IoT services ing, making it very difficult to agree on an unified archi-
are typically composed of a set of distributed components, tecture. Recently, a few promising definitions of an unified
running in different locations and connected through dynamic Fog architecture have been proposed [5], [6], but they are too
networks. The emergence of these technologies led to explo- general, omit detailed descriptions, and do not address multiple
sive growth in data generation that needs to be processed with important factors, such as system scalability, interaction and
lowest possible latency. communication among the Fog devices, and resource mapping.

978-1-5386-6488-9/18/$31.00 ©2018 IEEE

The shared complexity of the Fog and the influence of the present an intriguing early-stage concept for Fog orchestra-
recent IoT trends moving towards deploying and intercon- tion and architecture management by implementing genetic-
necting extremely large sets of pervasive devices and sensors, based heuristics, focused on gradual optimization of the IoT’s
aggravates this issue even more. The requirements for defining workflows in relation to the QoS requirements. To achieve
a unified Fog architecture are very high, as we deal with a balance between the accuracy and time efficiency, the authors
complex distributed structure capable of processing a multi- separate the computation among multiple workers and utilize
tude of parallel heterogeneous tasks. Adding the architectural a centralized master node for result aggregation and decision
heterogeneity of the various computational devices utilized making. In addition to inducing a single point of failure and
in the Fog, it is increasingly acknowledged that novel Fog computational bottleneck, this approach can lead to identifica-
architectural approaches capable of adapting and scaling in tion of false global extremes, which influences the quality and
response to the external environment and distributed appli- accuracy of the results. Furthermore, in [10] a novel concept of
cations need to be explored. Therefore, we strongly believe Osmotic computing, which is a new paradigm for supporting
that the highly parallel nature of the Fog can be described an efficient execution of IoT services and applications at the
through an analogy with one the of most powerful and efficient network edge has been presented. This intriguing approach for
processing system, the human brain, referring to its ability to Edge/Cloud computing presents interesting concepts, which
adapt its own structure and functions following the changes in could be utilized as an extension or alternative to the Fog.
the environment, a property called plasticity. A promising orchestration model for Fog resources pro-
In this paper we introduce a promising new nature-inspired visioning based on a service-oriented approach is proposed
Fog architecture, named SmartFog, capable of providing low in [11], enabling container-based resource provisioning in
decision making latency and adaptive architecture structuring. distributed architectures through a hybrid orchestration ar-
By utilizing novel algorithms and techniques from the fields chitecture. Unfortunately, this approach does not focus on
of multi-criteria decision making [7], graph theory [8] and extending the Fog architecture to consider non-functional
machine learning [9] we model the Fog as a distributed parameters and cannot adaptively react to the changes in
intelligent processing system, therefore emulating the function the workload. Similarly to the previously described concept,
of the human brain. In our analogy, the Fog devices are there is a possibility for a single point of failure if additional
modeled to mimic the function of the neurons, while the redundancy measures are not taken due to the centralized
synapses are correlated with the communication channels. architecture of the orchestration module.
The Fog devices are capable of self-clustering into multiple The work in [12] proposed a novel resource management
functional areas, optimized to support the functioning of a methodology for the Fog considering multiple different factors,
given IoT application and capable of restructuring in relation such as user behavior, Fog devices availability, and services
to the intensity and pattern of the sensory data flow. The price. Although this approach implements intriguing concepts
pervasive IoT devices and sensors are represented by the for resource estimation in fine-grained manner, it does not
sensory nervous system. For example, the temperature sensors explore methodologies adaptive architecture management.
can be related to the thermoreceptors, or the surveillance The work in [13] presents a conceptual framework for
cameras to the photoreceptors. The Cloud functionality takes resource provisioning based on a centralized Cloud-Fog mid-
a twofold role in our model: (i) it will provide backbone dleware together with a hierarchical Fog orchestration control
for supporting communication between the different functional system to manage the provisioning of the computational
areas, therefore emulating the function of the corpus callosum, resources in IoT and Fog environments on a local level.
and (ii) it will enable long-term storage of the processed data However, this approach does not consider any independent
sets. communication and self-adaptation between the orchestration
The paper is organized as follows. Section II summarizes control nodes, therefore limiting the possibility for deployment
the related research activities. Section III provides detailed of adaptive Fog architecture.
information on the related concepts and methodologies from III. BACKGROUND
the fields of multi-criteria decision making, graph theory, and
machine learning. Section IV explains the architectural model A. Multi-objective optimization
and formulates the novel methodologies. Section V provides In this work we extend and utilize essential concepts from
implementation details of the simulated Fog environment and the area of multi-criteria optimization and decision making
the implementation of the utilized methodologies. Section VI with a main goal to enable efficient communication between
presents experimental results and Section VII concludes the the IoT devices and the Fog/Cloud systems above. In the most
paper. general sense, optimization is a process of identifying one or
multiple solutions, which correspond to the extreme values of
II. R ELATED WORK two or more objective functions within given constraints set. In
the cases in which the optimization task utilizes only a single
Recently, promising research initiatives have been started objective function it results in a single optimal solution. More-
in the research community, focused towards solving issues over, the optimization can also consider multiple conflicting
related to the IoT and Fog computing. The authors in [4] objectives simultaneously. In those circumstances, the process
will usually result in a set of optimal trade-off solutions, supervised and unsupervised [17]. In supervised classification,
so-called Pareto front. The task of finding the optimal set the features’ mapping from a set of input data vectors is
of Pareto solutions in the form of Pareto front is known in classified to a finite set of discrete labeled classes and it is
the literature as a multi-objective optimization [14] [15]. The modeled in terms of some mathematical function. On the
Pareto front is an essential tool for decision support and prefer- other hand, in unsupervised classification, called clustering,
ence discovery, whose shape provides new insights and allows no labeled data-sets are available. The aim of the clustering is
scientists to explore the space of non-dominated solutions, to separate a finite unlabeled data-sets into a finite and discrete
possibly revealing regions of interest that are impossible to set of clusters. For the purpose of efficient clustering of the
see otherwise. Fog devices into functional groups we utilize unsupervised
In general, the multi-objective optimization problem in- clustering technique called Spectral clustering [9]. Spectral
volves two or more objective functions which have to be either clustering is a technique that make use of the eigenvalues
minimized or maximized. The problem of optimization can be [18] of the similarity matrix Sij = s(xi , xj ) of the data points
formulated as: xi , xj to perform dimensionality reduction before clustering
the data in relation to a reduced number of dimensions. The
min/max(f1 (Y ), f2 (Y ), . . . , fn (Y )) (1)
commonly used similarity measures in Spectral clustering are
where n ≥ 2 is the number of objectives functions f that we based on the Euclidean distance and the Gaussian kernel.
want to minimize or maximize, while Y = (y1 , y2 , . . . , yk ) is After performing dimensionality reduction, Spectral clustering
a region enclosing the set of feasible decision vectors. relies on more simple clustering techniques, such as k-means
[19], to group the similar data points in distinctive similarity
B. Betweenness centrality based sub-graphs. Spectral clustering is suitable for finding
In the research field of graph theory, the betweenness resemblance between different data points by utilizing the
centrality is used as a measure of node centrality within a concept of graph similarity, making it easily applicable for
graph, in relation to the number of shortest paths passing sub-diving of network graphs, such as the Fog environment.
through the node (vertex) [16]. For each pair of vertices
IV. A RCHITECTURAL MODEL
in a connected unweighted graph, there exists at least one
shortest path, such that the number of edges that the route The SmartFog architecture has been envisioned as a self
passes through can be minimized. Respectively, for weighted adapting system that, similarly to the human brain, can react
graphs the betweenness centrality utilizes the the sum of the to environmental changes. The SmartFog architecture aims
edges’ weights to find the shortest paths. The betweenness on providing a robust computational backbone to the IoT
centrality for each vertex is the number of the shortest paths platform underneath, while efficiently utilizing the Cloud
that pass through the vertex and can be represented through services above. The SmartFog architecture, depicted in Figure
the following equation: 1, is composed of three distinctive layers: (1) Cloud layer, (2)
Fog layer, and (3) IoT layer. SmartFog loosely integrates both
X σsd (n)
g(n) = (2) Cloud and IoT layers associated to the Fog, thereby enabling
σsd independent evolution and high degree of interaction.
s6=n6=d

where σsd is the total number of shortest path routes from

vertex s to vertex d, while σsd (n) represents the number of
shortest paths from vertex s to vertex d passing through vertex
n.
High value of the betweenness centrality usually implies
that a given vertex can reach others with lowest possible la-
tency. Moreover, high betweenness centrality can also indicate
that a given node lies on the path of many shortest routes.
Consequently, if one removes a node with large centrality it
can lengthen the paths between many other pairs of nodes.
This concept can be utilized in Fog computing in order to find
the most central Fog devices, therefore enabling more efficient
communication between the IoT devices and the Fog.
C. Spectral clustering
In the Big Data era the vital tool for dealing with large data-
sets is the concept of classification or grouping of data objects
into a set of categories or clusters. The classification of the Fig. 1. SmartFog architecture
objects is conducted based on the similarity or dissimilarity
of multiple features that describe them. Essentially, the clas- To provide a familiar environment to the IoT applications
sification methods can be divided into two categories, namely developers, SmartFog utilizes the Cloud as the main entry
point for application deployment and execution. The applica- ways is performed by the Cloud layer, which sorts the avail-
tion components, together with their functional requirements, able Fog devices based on dominance in multi-dimensional
need to be provided by the developers in a conventional search space. The approach that we take to perform the non-
manner through the Cloud layer. For example, the components domination sorting is based on the Fast Elitist Non-dominated
of a real-time video monitoring and analysis application can Sorting Genetic Algorithm (NSGA-II) [20]. In this approach,
be provided through the Cloud layer. Accordingly, the novel every Fog device from the full set of available devices, is
concepts, methodologies and technologies introduced in this checked with a partially filled set F 0 . To begin with, the
work are transparent for the IoT applications. first Fog device from the list of devices is initially kept in
The SmartFog architecture evolves around the Fog layer, the set F 0 . Afterwards, each Fog device f is compared with
which can potentially enable autonomous management and all members of the set F 0 one by one. If the Fog device f
provisioning of the computational and storage resources to the dominates any other fog device q in the set F 0 , then the device
IoT applications deployed through the Cloud layer. This layer q is removed from the set. On the other hand, if device f is
utilizes Spectral clustering techniques to perform efficient self- dominated by all member of F 0 , then f is ignored. Moreover,
clustering of the Fog devices, therefore emulating the function if device f is not dominated by any device in F 0 , then it is
of the human brain. The clustering process is autonomously included in F 0 . When all Fog devices known to the cloud
executed by a large set of specialized Fog devices called are checked, the remaining members of F 0 constitute the non-
communication gateways (depicted in yellow on Figure 1), dominated set.
that also act as main points for communication and decision The non-domination sorting algorithm is performed by
making between the Cloud and the Fog, thus emulating the considering the following conflicting criteria: (i) betweenness
function of the brain’s corpus callosum. For the selection of centrality, (ii) computational performance, and (iii) commu-
communication gateways or neurons, the Cloud layer utilizes nication latency to the Cloud. The main purpose of the non-
multi-criteria sorting and decision making algorithms by domination soring is to identify the Fog devices with highest
considering three distinctive criteria: betweenness centrality, number of available resources and the most optimal position
computational performance, and communication latency to the in the Fog system. The betweenness centrality, as one of the
Cloud. The utilization of the concept of betweenness centrality, objectives in the non-domination sorting algorithm, provides
as an objective in the multi-criteria sorting algorithm, enables information on how well the each Fog device is connected
identification of the topologically most central Fog devices in relation to the other Fog devices, the Cloud layer above
with sufficient amount of resources, capable of performing the and the IoT layer below. Respectively, the computational per-
self-clustering of the Fog devices. formance objective provides information on the computational
Afterwards, the selected communication gateways collabo- capacity of the Fog devices. For the purposes of the SmartFog
ratively cluster the remaining Fog devices, so-called neurons, system, the computational performance is represented through
into specific functional areas (depicted with colored dashed the number of instructions that the processor of the Fog
lines), thus organizing the Fog similar to a human brain. From device can execute in one second - MIPS. The communication
a strictly technical point of view, the functional areas are latency objective is represented through the virtual link latency
clusters, or more concretely connected graphs, of logically between a given Fog device and any other device. In our
grouped Fog neurons, capable of optimally supporting specific implementation, historical data on the previous data transfers
types of IoT applications. The clustering process of the func- is being preserved to calculate the average latency between
tional areas is conducted by considering multiple performance any two devices in the Fog environment.
related criteria. Therefore, each functional area is optimized The process of non-domination sorting algorithm results
for a specific application type, including computation, memory with a set of multiple optimal solutions, visualized in the
and network-intensive. The clustering process is continuously form of Pareto front. As all of the Fog devices represented
conducted in an adaptive manner that reacts to changes in the in the Pareto front are optimal, we utilize automated decision
environment, such as unforeseen changes in the architecture making strategy, which considers a-priory input information
(addition or removal of Fog devices). To enable fast clustering to select only the required number of Fog devices to act
in distributed environments, we utilize the Spectral clustering as Communication gateways. The decision making algorithm
technique, described in Section III. considers the number Nf and the type F of the functional
areas to give priority which Fog devices to be selected. For
A. Communication gateways selection example, if a given functional area should be optimized for
The process of Communication gateways selection is of processing, then the decision making algorithm will select the
utmost importance for reducing the decision making latency Fog device with the highest computational performance from
and increasing the resources utilization. Essentially, the Com- the set F 0 of Pareto optimal Fog devices. The selection process
munication gateways act as a high-bandwidth aggregation of the Communication gateways is presented in Algorithm 1.
points between the edge and the Cloud. Furthermore, they
perform essential operations for clustering of the Fog devices B. Functional areas clustering
in functional areas and mapping of the IoT applications to the The concept of functional areas, introduced by the Smart-
physical resources. The selection of the Communication gate- Fog architecture, evolves around the notion of Fog devices’
Algorithm 1 Communication gateways selection multi-criteria non-domination sorting and spectral clustering
Input: algorithms.
N . Number of Fog devices
A. Simulation environment
L = (L1 , L2 , ..., LN ) . List of Fog devices
C = (C1 , C2 , ..., CN ) . CPUs per Fog device We simulated an unstructured Fog overlay network system
M = (M1 , M2 , ..., MN ) . Memory per Fog device on top of the iFogSim [21], which is an efficient toolkit for
T . Topology of the Fog layer modeling and simulating resource management techniques in
Nf . Number of functional areas IoT and Fog computing environments. To accommodate the
F = (F1 , F2 , ..., FNf ) . Type of the functional area novel concepts introduced by SmartFog we have extended
1: while i < N do iFogSim to support the automated gateways selection and
2: Bi ← evaluate betweenness centrality(Li , T ) devices’ clustering.
3: Ei ← evaluate F og device(Ci , Mi , Bi ) Within the simulation environment we assume that every
4: i←i+1 Fog node in the overlay network have a specific processing
5: S ← non domination sorting(E) capacity and system architecture, such as ARM or x86.
6: D ← automated decision making(S, Nf , F ) Furthermore, every node has limited operating memory and
storage disk capacity with a fixed number of stored data items
of varying size. In addition a specific scheme has been defined
to allow every Fog node to leave and join the network after a
grouping into distinctive clusters optimized for special types
random time interval, which is an important scenario supported
of IoT applications, such as compute or memory intensive. Es-
by the SmartFog architecture.
sentially, the clustering process enables grouping of logically
similar Fog devices by considering various resource related B. Multi-criteria decision making and spectral clustering
criteria. We perform the gateway selection based on multiple con-
The clustering process and the formation of the functional flicting objectives, which have been modeled as a part of the
areas is performed in distributed manner among the commu- information collected from the initial Fog overlay simulation.
nication gateways. Every communication gateway performs We utilize modified NSGA-II multi-objective optimization
Spectral clustering by considering the available computing algorithm to perform the non-domination sorting of the Fog
and memory resources of the available Fog devices. A-priory devices. In addition to this, we also utilize low-latency auto-
input from the Cloud layer is used to steer the clustering mated decision making algorithm. To instantiate NSGA-II as
process in the preferred direction, therefore creating clusters of a main component of the communication gateway selection
Fog devices with similar amount of available processing and module, we have extended the jMetal [22] object-oriented
memory resources. For example, if a communication gateway Java framework for multi-objective optimization problems.
is intended to create compute optimized functional area, the We have implemented particular modifications in the jMetal
Spectral clustering will identify all similar Fog devices, which framework to deal with the specific characteristics of the
have sufficient amount of processing resources. The process proposed architecture. More concretely, we have modified the
of functional areas formation is presented in Algorithm 2. It non-domination sorting algorithm, as the standard operators
is essential to be noted, that every communication gateway of the jMetal framework do not support independent multi-
creates each own cluster, therefore a given Fog device can be criteria sorting. Furthermore, jMetal was extended to support
member of multiple functional areas, provided it has sufficient automated decision-making by utilizing the algorithm pro-
resources. posed in [23]. This algorithm performs low-latency decision
Algorithm 2 Functional areas clustering making by dividing the Pareto front in multiple regions based
on the a-priory knowledge that gives priority to specific
Input:
criteria. Therefore, this algorithm is suitable for the SmartFog
N . Number of Fog devices
architecture. Moreover, we modified the jMetal framework
k . Number of clusters
and extended its API to support integration with the iFogSim
C = (C1 , C2 , ..., CN ) . CPUs per Fog device
simulation environment.
M = (M1 , M2 , ..., MN ) . Memory per Fog device
Lastly, the Spectral clustering was implemented by extend-
G . Value of the Gaussian filter
ing the WEKA data mining framework [24]. The implemen-
1: L ← create dense matrix(C, M )
tation of the spectral clustering was based on the algorithm
2: Y ← f ind eigen vectors(L, k)
proposed in [25]. Additionally, the Spectral clustering im-
3: Km ← apply k means(Y, k)
plementation was integrated within the jMetal framework,
4: Fk ← create f unctionalc lusters(Km )
thus fully supporting the novel concepts introduced in the
SmartFog.
V. I MPLEMENTATION VI. E XPERIMENTAL EVALUATION
In this section, we discuss the essential implementation In this section we present broad experimental evaluation
details of the proposed Fog architecture with respect to the of the proposed concepts for Fog architecture management.
We conducted the evaluation based on a set of indicators, As it can be observed from the simulation results, the
specifically selected for the analysis of the non-domination execution time of the betweenness centrality calculation rises
sorting and clustering operations, covering: (i) the communi- linearly with the size of the network overlay. Nevertheless,
cation gateway selection, and (ii) the functional area grouping. even for 40 Fog devices, which can potentially serve thousands
Furthermore, we conducted extensive simulation-based evalu- of Edge and IoT devices, the execution time is below 40 ms.
ation to investigate the influence of the introduced concepts in Furthermore, on Figure 3, we provide evaluation data of the
the overall efficiency of the Fog architectures. non-domination sorting algorithm. The experimental scenario
is identical with the one described above. The execution
A. Communication gateway non-domination sorting and de- time presented in the Figure includes the time required for
cision making the decision making to be performed, while omitting the
The essential feature of the SmartFog architecture is the betweenness centrality calculation. We can observe that for
introduction of the communication gateways, which act both our testbed, the non-domination sorting was very efficient and
as a aggregation points between the Edge and the Cloud, and induced latencies below 5 ms.
are actively involved in the creation of the functional areas.
For those reasons, it is important to evaluate the behavior,
efficiency and scalability of the implemented non-domination
sorting and decision making algorithms. For the evaluation
purposes we utilize network overlays with sizes varying from
20 to 40 vertices, i.e. Fog devices. In the simulation test-bed
we assumed that the Fog devices have limited computation
resources, ranging from 800 to 1200 MIPS. In relation to the
memory and storage resources, we assume that each device
can have between 1 and 4 GB of operating memory and small
amount of solid-state based storage. All experiments were
repeated 100 times and the median value, together with the
standard deviation, are presented in the Figures below.
Initially, the evaluation activities were focused towards
validation of the concepts, introduced from the areas of graph
theory and multi-criteria decision making. Therefore, we first Fig. 3. Computation scalability of the Non-domination sorting algorithm
explore the connectivity characteristics, expressed as between-
ness centrality, of every Fog device, which are utilized in the
B. Functional areas clustering
multi-criteria non-domination sorting. The exploration of the
scalability potential of the betweenness centrality algorithm is The other concept, exploited by the SmartFog architecture,
essential, as it should be executed each time there are changes is the Spectral clustering algorithm, which is being utilized by
in the highly volatile IoT/Fog network overlay. Moreover, this the communication gateways to create specialized functional
process requires calculation of all shortest path routes in a areas. The current implementation of the SmartFog architec-
given topology, therefore increasing the complexity of the ture supports the creation of computational and memory opti-
algorithm. The evaluation of the the betweenness centrality mized functional areas. For the purpose of Spectral clustering
algorithm is presented on Figure 2. we again consider varying sizes of the network, which ranges
from 20 to 40 Fog devices. The evaluation results, presented
on Figure 4, clearly show that the algorithm scales very well,
with latencies in the range of 300 ms and very low standard
deviation values.
C. Simulation experiment
In order to explore the influence of the introduced con-
cepts on the overall efficiency of the Fog architectures we
have conducted extensive simulation-based analysis. We have
simulated a Sense-Process-Actuate Model (SPAM) with Edge-
ward placement strategy. In SPAM, the sensors continuously
emit data in the form of tuples. Tuple is basic unit of
communication between different entities in iFogSim, which
is specified by the following parameters: data source, data
destination and amount of required computation and network
resources. These tuples are transmitted to the Fog devices
Fig. 2. Computation scalability of the betweenness centrality algorithm as data streams. The IoT applications modules, executed on
 Fog to the Cloud. The simulation results from the loop delay
evaluation are presented on Figures 5 and 6.

Fig. 4. Computation scalability of the Spectral clustering algorithm

the Fog devices, process this data and transmit it back to

Fig. 5. SPA loop latency in Fog environment
the actuators for performing a specific action. For the Edge-
ward placement strategy, specific application’s modules are
deployed at edge of the network, while the remaining modules The evaluation results, both for the SPA and the PC loop
are placed to act as communication gateways to the Cloud. latency, clearly show that SmartFog can reduce the decision
Therefore part of the data is being processed by Fog devices making latency by up to 8%, which is significant improvement
placed at edge of network and the remaining data is transmitted in volatile environments such as the Fog. In the cases when
to the Cloud for further processing. the data needs to be processed in the Cloud, more concretely
within the PC loop, we can conclude that for smaller overlays
In our simulation model, the sensors and actuators are
SmartFog can reduce the communication latency, while for
attached to the edge devices (mobiles). These mobiles are con-
larges overlays the latency remains the same. Furthermore, in
nected with the Fog devices through edge-gateways. Whereas
the case of the SPA loop, SmartFog can significantly reduce
on the other end, the Fog devices are connected with the Cloud
the latency, both for small and large overlays.
through the communication gateways. We have used edge-
ward placement strategy in our simulation. Therefore we have
defined two process-control loops to calculate the end-to-end
latency or IoT application loop delay. The first is a sense-
process-actuate (SPA) loop in which the Fog devices receive
the data from the sensors, and then process it and transmit it
back to the actuators for further action. The second one is the
process-control (PC) loop in which, through communication
gateways, the Cloud receives data from the Fog devices, and
then process it and transmits it back to the source Fog devices.
Therefore, we conducted a series of experiments to evaluate
the performance of the SmartFog against unoptimized-Fog in
relation to the SPA and PC loop delays. These series consisted
of three experimental scenarios with three overlay topologies
with sizes of 20, 30, and 40 Fog devices respectively. The
Fog devices in the overlays are connected randomly with one
another creating a mesh kind of topology. Furthermore, the
Fog devices are connected with the sensors and actuators, Fig. 6. PC loop latency in Fog environment
which are continuously emitting tuples. The Fog devices are
also attached to Cloud through the communication gateways. Moreover, we have further focused our experimental evalu-
For calculating the SPA loop delay, we have used tuples with a ation on the network load induced by the IoT applications. For
requirement of 1000-8000 MIPS and 100 bytes network width. this reason, we have evaluated the total network traffic over
Moreover, for the PC loop delay, the tuples have minimal the overlays for the SmartFog architecture and unoptimized
requirement of 40000 MIPS and 100 bytes. In our experiments Fog. The results, presented on Figure 7, clearly show that for
the tuples are sent randomly from the sensors to the Fog all network sizes the communication load can be reduced by
devices, while SmartFog manages the data transfers from the up to 13%.
 [6] Flavio Bonomi, Rodolfo Milito, Preethi Natarajan, and Jiang Zhu. Fog
computing: A platform for internet of things and analytics. In Big Data
and Internet of Things: A Roadmap for Smart Environments, pages 169–
186. Springer, 2014.
[7] Dragi Kimovski, Julio Ortega, Andrés Ortiz, and Raúl Baños. Parallel al-
ternatives for evolutionary multi-objective optimization in unsupervised
feature selection. Expert Systems with Applications, 42(9):4239–4252,
2015.
[8] Ulrik Brandes, Stephen P Borgatti, and Linton C Freeman. Maintaining
the duality of closeness and betweenness centrality. Social Networks,
44:153–159, 2016.
[9] Ulrike Von Luxburg. A tutorial on spectral clustering. Statistics and
computing, 17(4):395–416, 2007.
[10] Massimo Villari, Maria Fazio, Schahram Dustdar, Omer Rana, and Rajiv
Ranjan. Osmotic computing: A new paradigm for edge/cloud integration.
IEEE Cloud Computing, 3(6):76–83, 2016.
[11] Mathias Santos de Brito, Saiful Hoque, Thomas Magedanz, Ronald
Steinke, Alexander Willner, Daniel Nehls, Oliver Keils, and Florian
Schreiner. A service orchestration architecture for fog-enabled infras-
tructures. In Fog and Mobile Edge Computing (FMEC), 2017 Second
Fig. 7. Network load in Fog environment
International Conference on, pages 127–132. IEEE, 2017.
[12] Mohammad Aazam and Eui-Nam Huh. Dynamic resource provisioning
through fog micro datacenter. In Pervasive Computing and Com-
VII. C ONCLUSION munication Workshops (PerCom Workshops), 2015 IEEE International
Conference on, pages 105–110. IEEE, 2015.
In this paper we introduce a promising novel nature-inspired [13] Olena Skarlat, Stefan Schulte, Michael Borkowski, and Philipp Leitner.
Fog architecture, named SmartFog, capable of providing low Resource provisioning for iot services in the fog. In Service-Oriented
Computing and Applications (SOCA), 2016 IEEE 9th International
decision making latency and adaptive architecture manage- Conference on, pages 32–39. IEEE, 2016.
ment. By utilizing novel algorithms and techniques from the [14] Jürgen Branke, Kalyanmoy Deb, and Kaisa Miettinen. Multiobjective
fields of multi-criteria decision making, graph theory and optimization: Interactive and evolutionary approaches, volume 5252.
Springer Science & Business Media, 2008.
machine learning we model the Fog as a distributed intelligent [15] Dragi Kimovski, Nishant Saurabh, Sandi Gec, Vlado Stankovski, and
processing system, therefore emulating the function of the Radu Prodan. Multi-objective optimization framework for vmi distribu-
human brain. One of the key strengths of our approach are tion in federated cloud repositories. In European Conference on Parallel
Processing, pages 236–247. Springer, 2016.
the ability (i) to identify the most optimal aggregation points [16] Marc Barthelemy. Betweenness centrality in large complex networks.
between the Cloud and the Edge, and (ii) to adaptively group The European Physical Journal B-Condensed Matter and Complex
the Fog devices in optimized clusters based on similarity. We Systems, 38(2):163–168, 2004.
[17] Rui Xu and Donald Wunsch. Survey of clustering algorithms. IEEE
have implemented, integrated and evaluated the introduced Transactions on neural networks, 16(3):645–678, 2005.
concepts as essentials elements of the SmartFog environment. [18] Alan Edelman. Eigenvalues and condition numbers of random matrices.
Based on the evaluation results, it can be concluded that the SIAM Journal on Matrix Analysis and Applications, 9(4):543–560, 1988.
[19] Tapas Kanungo, David M Mount, Nathan S Netanyahu, Christine D
proposed nature-inspired Fog architecture can provide up to Piatko, Ruth Silverman, and Angela Y Wu. An efficient k-means
8% latency reduction, and 13% reduction in network load. clustering algorithm: Analysis and implementation. IEEE transactions
Regarding the future research activities, we plan to extended on pattern analysis and machine intelligence, 24(7):881–892, 2002.
[20] Kalyanmoy Deb, Samir Agrawal, Amrit Pratap, and Tanaka Meyarivan.
the current environment to support proximity aware data A fast elitist non-dominated sorting genetic algorithm for multi-objective
management and adaptive resource provisioning. optimization: Nsga-ii. In International Conference on Parallel Problem
Solving From Nature, pages 849–858. Springer, 2000.
ACKNOWLEDGMENT [21] Harshit Gupta, Amir Vahid Dastjerdi, Soumya K Ghosh, and Rajkumar
Buyya. ifogsim: A toolkit for modeling and simulation of resource man-
This work is being accomplished as a part of the Tiroler agement techniques in the internet of things, edge and fog computing
Cloud Project: Energiebewusste Föderierte Cloud für Anwen- environments. Software: Practice and Experience, 47(9):1275–1296,
dungen aus Industrie und Forschung, funded by the Bridge 2017.
[22] Juan J Durillo and Antonio J Nebro. jmetal: A java framework for multi-
programme under grant agreement No 848448. objective optimization. Advances in Engineering Software, 42(10):760–
771, 2011.
R EFERENCES [23] Roland Matha Dragi Kimovski, Sasko Ristov and Radu Prodan. Multi-
[1] Mahadev Satyanarayanan. The emergence of edge computing. Com- objective service oriented network provisioning in ultra-scale systems.
puter, 50(1):30–39, 2017. In EUROPAR Workshops 2017. Springer, 2017.
[2] Amir Vahid Dastjerdi, Harshit Gupta, Rodrigo N Calheiros, Soumya K [24] Mark Hall, Eibe Frank, Geoffrey Holmes, Bernhard Pfahringer, Peter
Ghosh, and Rajkumar Buyya. Fog computing: Principles, architectures, Reutemann, and Ian H Witten. The weka data mining software: an
and applications. arXiv preprint arXiv:1601.02752, 2016. update. ACM SIGKDD explorations newsletter, 11(1):10–18, 2009.
[3] Prateeksha Varshney and Yogesh Simmhan. Demystifying fog comput- [25] Andrew Y Ng, Michael I Jordan, and Yair Weiss. On spectral clustering:
ing: Characterizing architectures, applications and abstractions. arXiv Analysis and an algorithm. In Advances in neural information processing
preprint arXiv:1702.06331, 2017. systems, pages 849–856, 2002.
[4] Zhenyu Wen, Renyu Yang, Peter Garraghan, Tao Lin, Jie Xu, and
Michael Rovatsos. Fog orchestration for internet of things services.
IEEE Internet Computing, 21(2):16–24, 2017.
[5] Luis M Vaquero and Luis Rodero-Merino. Finding your way in the fog:
Towards a comprehensive definition of fog computing. ACM SIGCOMM
Computer Communication Review, 44(5):27–32, 2014.