Interdependent Multi-layer Spatial Temporal-based Caching in
Heterogeneous Mobile Edge and Fog Networks
Vu San Ha Huynh
a
and Milena Radenkovic
b
School of Computer Science, University of Nottingham, Nottingham, U.K.
Keywords: Content Caching, Mobile Edge and Fog Networks, Network Multilayer Interplay, Spatial-temporal Locality.
Abstract: Applications and services hosted in the mobile edge/fog networks today (e.g., augmented reality, self-driving,
and various cognitive applications) may suffer from limited network coverage and localized congestion due
to dynamic mobility of users and surge of traffic demand. Mobile opportunistic caching at the edges is
expected to be an effective solution for bringing content closer and improve the quality of service for mobile
users. To fully exploit the edge/fog resources, the most popular contents should be identified and cached.
Emerging research has shown significant importance of predicting content traffic patterns related to users’
mobility over time and locations which is a complex question and still not well-understood. This paper tackles
this challenge by proposing K-order Markov chain-based fully-distributed multi-layer complex analytics and
heuristics to predict the future trends of content traffic. More specifically, we propose the multilayer real-time
predictive analytics based on historical temporal information (frequency, recency, betweenness) and spatial
information (dynamic clustering, similarity, tie-strength) of the contents and the mobility patterns of contents’
subscribers. This enables better responsiveness to the rising of newly high popular contents and fading out of
older contents over time and locations. We extensively evaluate our proposal against benchmark (TLRU) and
competitive protocols (SocialCache, OCPCP, LocationCache) across a range of metrics over two vastly
different complex temporal network topologies: random networks and scale-free networks (i.e. real
connectivity Infocom traces) and use Foursquare dataset as a realistic content request patterns. We show that
our caching framework consistently outperforms the state-of-the-art algorithms in the face of dynamically
changing topologies and content workloads as well as dynamic resource availability.
1 INTRODUCTION
User mobile devices today are increasingly intelligent
which leads to the explosive development of new
applications involving distributed real-time mobile
processing and increasing traffic demands (e.g. HD
video streaming, remote health care, critical
applications for public safety communications,
augmented/virtual reality apps and automatic
driving/traffic control). Varying mobility patterns,
network topology changes, potential disconnections
and resource restrictions in mobile environments pose
many challenges for the design and implementation
of future mobile network algorithms, particularly
content caching with an aim to bring contents
proactively as close as possible to the users and
improve the reliability and efficiency of mobile
edge/fog networks and users services. Typical
a
https://orcid.org/0000-0002-5472-5328
b
https://orcid.org/0000-0003-4000-6143
edge/fog networks consist of heterogeneous nodes
which can include end users and edge devices with
different computing resources and communication
capabilities (Liu et al., 2018). Our architecture design
assumes that the communication between edge/fog
nodes is handled in mobile opportunistic multi-hop
manner.
Existing opportunistic caching policies in mobile
edge/fog networks such as (Wang et al., 2014; Fricker
et al., 2012) typically cannot capture, predict and
adapt to the spatial-temporal locality of content
requests needed for more accurate content popularity-
based caching decisions because they rely on
assumptions that content interests occur
independently of users’ mobility and resources. More
specifically, when content caching predictions are not
sufficiently fine-grained, they result in increased
cache miss (i.e. the content request needs to traverse
34
Huynh, V. and Radenkovic, M.
Interdependent Multi-layer Spatial Temporal-based Caching in Heterogeneous Mobile Edge and Fog Networks.
DOI: 10.5220/0008167900340045
In Proceedings of the 9th International Conference on Pervasive and Embedded Computing and Communication Systems (PECCS 2019), pages 34-45
ISBN: 978-989-758-385-8
Copyright
c
2019 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
from subscribe to publisher), delay and resource
consumption. In order to tackle this complex
challenge, we propose a next-generation content
caching approach that is able to capture more
accurately the dynamic spatial-temporal correlation
of mobility and traffic patterns as well as the
mobility-content traffic interplay needed to support
more reliable, adaptive caching algorithms. Previous
research (Wang et al., 2017; Radenkovic et al., 2018)
has shown that the centralised solution is not scalable
in mobile complex heterogeneous network topologies
due to its high complexity and single point of failure
problem. Moreover, distributed solution technique
(Wang et al., 2017) may still cause high connectivity
overheads although it provides a cheaper computable
lower bound compared to the centralised solution. In
addition, previous research has also shown that
collaborative caching usually outperforms both
locally and centrally optimized algorithms (Saha et
al., 2013). Therefore, we propose a fully-distributed
predictive analytic and heuristic-based caching
approach which comprises of multi-layer
complementary real-time distributed predictive
heuristics to maintain the best possible trade-off
between caching performance and resource
consumption. Note that our focus is not to build a
protocol that forces nodes to cooperate to achieve
mutual benefit, but rather to design an underlying
algorithm that ensures no node attains lower utility by
collaborating with others, similarly to (Radenkovic
and Huynh, 2017; Wang et al., 2017; Radenkovic et
al., 2018). Existing research utilizes different
approaches to deal with trend prediction such as
reinforcement learning, Bayesian learning, Markov
chain (Ruan et al., 2019), or Exponential Moving
Average (Radenkovic and Huynh, 2017; Radenkovic
et al., 2018; Huynh and Radenkovic, 2018). Due to
the continuous nature of users’ mobility and content
requests, in this paper, we propose to apply the
concept of high-order Markov chain to our complex
real-time analytics to more accurately predict the
content traffic trends based on the historical
information of content traffics related users’ mobility.
This paper extends the multilayer
multidimensional predictive heuristics integrated in
CafRepCache (Radenkovic and Huynh, 2017;
Radenkovic et al., 2018) which is an predictive
adaptive collaborative cognitive forwarding and
caching framework in heterogeneous opportunistic
mobile networks. We propose a novel prediction
model that allows improved congruency with both the
underlying network and users demands, more
accurately capture the temporal and spatial locality of
mobility and content requests as well as mobility-
content traffic patterns interplay. We build on two
integral complementary multidimensional predictive
analytics and heuristics: i) temporal predictive
analytics and heuristics that captures the temporal
locality of content requests upon request access time,
enables more responsiveness to the rising trend of
newly high popular contents and fading out of older
contents over time as well as avoid one-timer contents
and mitigate flash crowd effect; ii) spatial predictive
analytics and heuristics that capture the spatial
locality of user mobility and content requests as well
as balance the trade-off between serving different
regions of contents’ subscribers. We exploit mobility-
content traffic patterns interplay and do not focus on
resource analytics (Radenkovic and Huynh, 2017;
Radenkovic et al., 2018) in this paper.
The paper begins by providing an overview of the
related work in section 2, section 3 describes and
discusses our models and multilayer novel predictive
heuristics, section 4 evaluates the effect of each
complement heuristic on caching performance in
mobile DTN across a range of metrics over two
different network topologies and a real-world
location-based service dataset for content workloads.
Section 5 gives a conclusion.
2 RELATED WORK
Authors in (Le et al., 2015) propose a forwarding and
cache replacement policy for SocialCache based on
content popularity driven by frequency and freshness
of content requests. As part of its replacement policy,
SocialCache may remove a cached content from the
network, thus reduce the cache hit ratio and increase
delays. This problem is exacerbated when the
resource is limited and the replacement rate is high as
(Le et al., 2015) is not resource and congestion aware.
Authors in (Zhang et al., 2014) propose Optimal
Cache Placement Based on Content Popularity
(OCPCP) that takes a caching decision based on the
frequency of content requests at a caching node. That
is, the more frequent requests for content, the higher
chance that content will be requested again. Time
Aware Least Recent Used (TLRU) (Bilal and Kang,
2014) is an extension of the simple LRU in which the
time stamp of an arriving content is recorded locally
by a single node. The arriving content is cached if the
average request time is smaller than the time stamps
of the stored contents. Authors in (Mardham et al.,
2018) propose Location-based Caching that uses
decay function measured by the function of distances
between subscribers and caching points and some
varying attributes, such as time or number of requests
Interdependent Multi-layer Spatial Temporal-based Caching in Heterogeneous Mobile Edge and Fog Networks
35
to classify contents and replace them when cache
memory is full. (Mardham et al., 2018) considers the
fairness problem that some contents belonging to
some specific location may be more important and is
cached often across the network, even if it is not very
popular. Authors in (Flores et al., 2017) proposed a
social-aware hybrid offloading strategy for load
balancing and computation sharing based on node's
stability which is measured by contact frequency and
duration in order to improve the availability of
offloading support for mobile users. However,
(Flores et al., 2017) does not support high topology
dynamics with intermittent disconnections and
dynamically changing publishers and subscribers
workload patterns.
Existing opportunistic caching policies such as
(Wang et al., 2014; Fricker et al., 2012) rely on an
assumption that the content distribution in the
networks approximately follows Zipf's law (Yoneki
et al., 2008; Breslau et al., 1999) or the
characterisation of content requests have been based
on Independent Reference Model (IRM) which
assumes that content requests occur independently.
However, authors in (Dabirmoghaddam et al., 2014;
Dán and Carlsson, 2010) have questioned the validity
of the IRM model and Zipf’s law model. According
to the proposal in (Dabirmoghaddam et al., 2014),
content requests often exhibit both temporal locality
and spatial locality. Researches on today’s social
networks (Dabirmoghaddam et al., 2014; D’Silva et
al., 2018) suggest that if content is requested at a
certain period in time, more likely it will be requested
again in the near future. In fact, the content is not
requested scattered randomly and independently over
time; but rather particular contents are interested at a
certain time interval, while its popularity gradually
fades out. Spatial locality of content requests is based
on the fact that content requests of the same content
are more likely to be issued by geographically close
areas. More precisely, the requests coming from a
specific region in space are more likely to be similar
than those collected over regions far apart (D’Silva et
al., 2018).
3 FULLY-DISTRIBUTED
PREDICTIVE MOBILE EDGE
AND FOG CACHING
In this section, we briefly describe CafRepCache
(Radenkovic and Huynh, 2017; Radenkovic et al.,
2018) framework, then we propose to extend its
integral multilayer fully-distributed predictive
heuristics.
CafRepCache (Radenkovic and Huynh, 2017;
Radenkovic et al., 2018) is a multi-path content and
interest forwarding and replication with adaptive
collaborative caching framework in heterogeneous
opportunistic mobile networks. It utilises fully-
localised and ego networks multi-layer predictive
heuristics about dynamically changing topology,
resources and content popularity to manage dynamic
trade-offs between minimizing the end-to-end latency
and maximising content delivery while enabling
resource efficiency and congestion avoidance.
CafRepCache relies on three theory foundation:
network science, social science and information
science as shown in Fig. 1.
Figure 1: CafRepCache theory foundation.
CafRepCache system is modelled as a network G
that consists of a set N of nodes
(
) and a set
E of edges, G = (N, E). As the connectivity of the
network and the state of the nodes change over time,
each of these sets is modelled as time series, thus N =
{
: t T} and E = {
: t T}. We denote with
a set of content files that can be requested by the
network. Each content
(or
for simplicity)
is published at time t. Node
may act as either
a subscriber of content
, denoted by

or a
publisher

or a caching point

at any time t
from any location while being mobile. Thus for any
content
, a set of subscribers who are interested in
is denoted as



and so on. Each
node in the network is able to perform predictive
analytics of multivariate mixed data (e.g. content and
mobility) as well as collaborate and exchange its local
observations with other neighbour nodes when two
nodes are in contact in order to capture and detect
PECCS 2019 - 9th International Conference on Pervasive and Embedded Computing and Communication Systems
36
various events (e.g. user connectivity patterns,
request patterns) in more accurate and responsive
manner without the need of global knowledge. We
define “ego network” of each node
: 
as a
network consisting of
together with the nodes they
are connected most recently (i.e. k-hops neighbours)
or nodes it meets at regular interval (most frequently)
over the time duration ∆T and all the links among
those nodes. In this way, ego network allows each
node to give its own regional or temporal perspective
of the network (or both are included).
3.1 Dynamic Complex
Temporal-Graph Heterogeneous
Network Topologies
In order to provide insights of users’ content traffic-
related mobility, connectivity patterns, this section
analyses CafRepCache in different underlying
network topologies with different degree of mobility,
connectivity patterns which helps us to better design
novel modelling analytics to capture and predict the
spatial-temporal correlations of content request
patterns with underlying network topologies.
Empirical evidence shows fixed and wired networks
are mostly scale-free, whereas mobile and
opportunistic networks can be either random or scale-
free. Thus, we analyse theoretically CafRepCache in
extremely heterogeneous random topologies as well
as scale-free topologies to understand fundamental
performance limitations of CafRepCache in different
realistic networks.
3.1.1 Complex Temporal Random Network
Topologies
In random networks, the majority of previous studies
in mobile networks assume Poisson contact processes
and model user encounters as independent Poisson
processes with rate λ due to the time between two
consecutive contacts of a pair of nodes follows
exponential distribution (Bornholdt and Schuster,
2006). The probability of some node
connecting with some other node
(Bornholdt and
Schuster, 2006) is:



.
The probability that node
connects with
exactly n other nodes (Bornholdt and Schuster, 2006)
within time T: P(n encounters)


. Let 
be the probability that content
is stored in the
cache of an arbitrary nodes, then the probability of the
cache miss in random network is:
Prob(content
is not in the cache) *
Prob(interest request never reaches correct caching
points within time T)
=   
*
 

= (  
*


  
.
3.1.2 Complex Temporal Scale-free Network
Topologies
Complex temporal scale-free networks are
characterized by a highly heterogeneous degree
distribution, which follow a power-law distribution
(Yoneki et al., 2008). Although the network may
change significantly over time, the degrees of its
nodes obey the power-law model at any time (Yoneki
et al., 2008).
The probability P(n encounters) of a node in the
network goes for large values of n as:


where is the shape parameter of the power-law
distribution and represents the degree of the power-
tail.
Then the probability of a cache miss of content
in scale-free networks is (   
*

 

.
3.2 Spatial-temporal Analytics and
Heuristics for Content Caching
In this section, we describe two non-trivial predictive
analytics and heuristics (i.e. K-order Markov
chained-based temporal heuristics and spatial
clustering heuristics) that cover different dimensions
of the spatial-temporal dynamics and mobility-
content traffic interdependence problem. When
combined together, they allow forming dynamic
transient interest and data dissemination topologies
based on predictive analysis and commonalities
between their interests, caches and retrieval histories
as well as connectivity histories.
3.2.1 Dynamic K-order Markov
Chained-based Temporal Heuristics
Each node in the network resolves the request
frequency, recency and betweenness in fully-
localised distributed manner. When two nodes are in
contact, they exchange their local observations and
continuously resolve the value of dynamically
changing predictive heuristic based on both its local
observation and the collaborative observations it gets
from others. We apply the concept of K-order Markov
chain on our complex analytics to predict the content
traffic trends based on the historical information of
Interdependent Multi-layer Spatial Temporal-based Caching in Heterogeneous Mobile Edge and Fog Networks
37
content traffics related users’ mobility. When K = 1,
a lot of historical state information is ignored and only
the information of the current moment is used. Such
limitations make the practical application of 1-order
Markov chain prediction method lack of prediction
accuracy. In line with (Ruan et al., 2019), compared
with other existing time-order-based prediction
methods, the K-order Markov chain performs much
better when the order number K = 2. We apply K-
order Markov chain to leverage efficiently historical
information of content requests and subscribers’
mobility to predict the future trends of content
traffics. The K-order is defined as:







We then introduce our method to measure the
temporal heuristics based on request frequency,
recency and betweenness as below.
Request frequency counters are additively
increased upon the arrival of a content request and
decreased through time. This is in order to ensure that
if the number of requests for a content has been
reduced, its popularity counter will be reduced
accordingly and the content will be subject to eviction
or offload to other nodes. Given a caching point
observes average f interests of content k during the
interval , the request frequency is measured as:


where is a
control parameter. Request frequency implies that the
more content requests have been observed by a
caching point during a short interval time , the
more likely that content will be requested in the same
interval, thus capture the temporal locality of content
requests
Request recency enables our caching design to
capture the content popularity trend in responsive
manner based on the recorded time stamp of recent
requests in different locations for each caching point,
thus allow to predict adaptively the emerging contents
that may become highly popular in near future and
contents that are currently considered as high popular
but will be less popular soon. Given a caching point n
observes the most recent interest of content k at

, then we denote

,

,

, etc. as the time that previous interests
have been recorded by the caching point. The request
recency is measured as how likely a recent content
request will trigger a subsequent request at the current
time.











Request recency analytic shows that the smaller
gap between current time and recent requests
observed in the past of content k, the higher chance
that content k will be requested soon in the future.
Request betweenness provides the trade-off
between current observed content popularity versus
long terms interest in it in order to balance between
potentially one-timer contents or fake news and long-
term useful content. The time gap between
continuous requests in the period of time  is
measured as:

 



,

 

,…
Then the request betweenness heuristic is
measured as:





The novel temporal heuristics are the combination
of request frequency, recency and betweenness which
allow CafRepCache to choose the best suitable
contents to cache and when to cache by predicting in
real time the locality trend of content request patterns
over time in different locations and avoid losing
valuable contents by reducing the caches for one-
timers contents and fake news.
 
 

.
3.2.2 Dynamic Spatial Clustering Heuristics
As the requests are more likely to be similar in a
specific region, we propose the fully-localised
distributed spatial heuristics that aim to classify,
recognize and predict content interests coming from
dynamic changing clusters of subscribers. The spatial
heuristic shows that the higher request rate coming
from the same localised region or dynamic cluster of
different subscribers, the higher likely that content
will be requested again by other subscribers within
that location. Given a caching point observes interest
requests of content k from a set of subscribers
within a time interval  , the spatial heuristic is
measured by the clustering coefficient (Nicosia et al.,
2013), similarity, closeness and tie strength
(Radenkovic and Huynh, 2017; Radenkovic and
Grundy, 2011; Daly and Haahr, 2007) between
different subscribers of the same content as below:
 

 
 

in which the similarity value (Daly and Haahr,
2007; Nicosia et al., 2013) between
within
a time interval ∆t is:





where 


is the similarity in contacts between two subscriber
of the content k. The closeness centrality (Daly
PECCS 2019 - 9th International Conference on Pervasive and Embedded Computing and Communication Systems
38
and Haahr, 2007; Nicosia et al., 2013) of
is a
measure of how close
is to any other node in
. It
is measured as the inverse of the average distance
from
to any other node in
:


where

is
the distance between
and
in the set of subscribers

.
The node tie strength value (Radenkovic and
Huynh, 2017; Radenkovic and Grundy, 2011; Daly
and Haahr, 2007) between of subscribers
within a time interval ∆t is:







where



measures the frequency of
contacts between
;


measures the
recency of contacts between
and


indicate the relative distance by hops between
.
The complex temporal graph metrics of contact
frequency, recency and topology distance allow
congruency with the underlying dynamic changing
network topology and connectivity. In order to
balance the trade-off between serving contents
requested from a specific local region of highly
connected subscribers and from multiple less-
connected subscribers, we favours the weak tie
strength which helps to give a wider and broader
long-term predictive content popularity instead of
only favouring and serving the contents requested
from a local region of highly connected subscribers.
Spatial heuristics allow CafRepCache to choose
the most suitable caching points to cache popular
contents based on its relative location with the
subscribers and between the subscribers themselves.
3.3 CafRepCache’s Combined Heuristics
In section 3.2, we described our novel approach for
opportunistic caching protocol that, we argue, is
essential to be able to capture the spatial-temporal
locality of mobility and content requests as well as the
mobility-content traffic interplay. The above two
non-trivial heuristics
cover different
dimensions of the mobility-content traffic pattern
interdependences problem and when combined they
allow managing a number of trade-offs we identified.
Each caching point
in the network resolves and
combines the two heuristics to measure

denoted as the popularity of
during the interval
time .
 implies the probability of how
likely the content
will be requested in a period of
time:



where
is the weighting factor of each heuristic,

is the respective utilities of each
heuristic as measurements of their relative gain, loss
or equality, calculated as pair-wise comparison
between the node’s own heuristics and that of the
encountered contacts:




 
The predictive analytics are resolved by caching
nodes’ local observations and collaborative
observations from neighbours within its ego network
in order to allow each caching point have a more
regional converged perspective of the network
without the need of global network knowledge.
The total content popularity heuristic is measured
as:






.
4 EXPERIMENT SETUP AND
EVALUATION
This section presents an evaluation of our caching
algorithm, first introducing a set of state-of-the-art
caching policies as competitive caching algorithms
and metrics for comparing the experimental results.
For the underlying network topology and mobility
patterns, we use a simulation-driven data trace with
two very different network topologies: random
network and real-world mobility traces Infocom
(Scott et al., 2006) in ONE simulator (Keränen et al.,
2009) as scale-free network in order to give a deeper
and more accurate performance overview of
CafRepCache. We use Foursquare New York Dataset
(Yang et al., 2014) as a real trace for content requests.
This dataset is collected through location-based
service Foursquare API
(https://developer.foursquare.com/) describing the
spatial-temporal locality of content requests in terms
of user interests at public venues, it contains 227,428
subscriptions of 18,201 users in different locations of
New York city during the period of 10 months. Each
check-in is associated with its time stamp, its GPS
coordinates and its semantic meaning (represented by
fine-grained venue-categories).
We compare and evaluate CafRepCache on the
overall caching performance measured by different
criteria (e.g. cache hit ratio, latency, eviction rate,
etc.) against multiple state-of-the-art and benchmark
proposals: SocialCache (Le et al., 2015), Optimal
Cache Placement Based Content Popularity (OCPCP)
(Zhang et al., 2014), Time Aware Least Recent Used
Interdependent Multi-layer Spatial Temporal-based Caching in Heterogeneous Mobile Edge and Fog Networks
39
Table 1: Values of the simulation parameters.
Parameter
Value
Complex temporal network
topologies
Random network, Scale-free
network (Infocom)
Content request pattern
Foursquare New York
Number of nodes
50-100
Simulation duration
1 - 3 hours
Request rate
1-25 request/min
Number of contents

- 
File size
1 MB - 8.4 MB
Interest packet size
8 kB - 128 kB


0.1 - 0.6%
(TLRU) (Bilal and Kang, 2014) and LocationCache
(Mardham et al., 2018). We have run six increments
of the number of subscribers and publisher ranging
from 10% to 60% of the total number of nodes. Due
to limited space, we report on experiments with
increasing number of subscribers, but note that the
results for increasing number of publishers are similar
to the ones presented here. Without losing generality,
we assume that 25% of node population are
publishers and varying the number of subscribers. All
experiments are repeated ten times and averaged with
different random subscribers and publishers. For each
experiment, either the cache hit ratio or average
latency (in hops) or eviction ratio will be shown. The
general simulation parameters details are shown in
Table 1.
Fig. 2 shows the spatial and temporal correlation
of content traffic (i.e. temporal requests pattern of
mobile subscribers) in Foursquare dataset for a
content in different locations over time.
Figure 2: Spatial-temporal correlation of content requests.
It shows the temporal patterns (similarity) of
content traffic during weekdays and weekend: if a
content is requested at a certain point in time, more
likely it will be requested again in near future. In fact,
nor are the content references scattered randomly and
independently over time; rather, a content might be of
particular interest at a certain time interval, while its
popularity gradually fades out. The locations of
mobile subscribers feature different degrees of
similarity in content request such that the location 1
and 2 which are relatively close to each other have
similar request patterns compared to that of location
3 which are far away from others. This captures the
impact that the geographical diversity of the users has
on the observed trace of requested contents by them.
More precisely, the requests coming from a specific
region in space are more likely to be similar than
those collected over regions far apart.
In order to understand the scalability of caching
points with regarding to the increasing number of
subscribers, we vary the number of subscribers to
evaluate the number of caching points in Fig.3,
average latency (measured by the number of hops) in
Fig. 4 and cache hit ratio (which refers to how many
interest packets are matched with the contents in
caching points without being forwarded to
publishers) in Fig.5 that indicate the efficiency of
caching decisions and locations in random topology
and scale-free topology.
Figure 3: Number of caching points vs. number of
subscribers.
Fig. 3 shows that the relative number of
CafRepCache caching points increases from 12% to
23% in random networks and from 2 to 11% in scale-
free networks regarding the growth of subscribers.
We argue that random networks with short average
paths and low clustering require more number of
caching points to serve the dynamic mobile
PECCS 2019 - 9th International Conference on Pervasive and Embedded Computing and Communication Systems
40
subscribers while scale-free networks with high
social character need less number of caching points
which converge to 11% regarding the increasing
number of subscribers from 40% to 60%. Fig. 4 and
5 shows that CafRepCache achieves 79.4% and
92.4% cache hit ratio while the hop count average
from the caching points to subscribers are 3.01 and
2.7 in random network and scale-free network
respectively.
We argue that in random network, CafRepCache
benefits from its cache redundancy mechanism that
select highly suitable locations for caching and
replication when needed as adaptive replication and
caching are both necessary to address multi-user data
communications in dynamic fragmented and sparse
topologies. In scale-free network, CafRepCache
utilities its multidimensional predictive analytics and
complex temporal graph metrics to make caching
decisions in a predictive manner and congruent with
the underlying network mobility, connectivity and
content interest.
Figure 4: Average hops count between caching points and
subscribers.
We show graphs of random network topologies as
a worst case scenario in Fig.3-5 and we will focus on
scale-free network topologies for the rest of our
experiments as it allows our caching decision
makings to leverage the spatial-temporal correlations
of mobility and traffic patterns as well as mobility-
traffic interdependence.
To evaluate the effectiveness of the content
request frequency heuristic, we vary the content
request patterns which follows different content
popularity skewness (0.6-1.1) utilized in Hawkess
process (Dabirmoghaddam et al., 2014) and measure
Figure 5: Success ratio vs. number of subscribers.
the cache hit ratio. In line with (Dabirmoghaddam et
al., 2014), when α becomes larger, there is small
number of contents that account for the majority of
total requests, or being requested more frequently
compared to the others.
Fig. 6 shows that our cognitive caching achieves
the highest cache hit ratio (ranging from 31% to
86.4%) compared to other competitive algorithms
(OCPCP, TLRU, SocialCache, LocationCache)
regarding the increase of popularity skewness α. We
observe that higher α leads to bigger gap between
CafRepCache and others. This is because the request
frequency heuristic takes advantage of highly skewed
content popularity and content request temporal
locality to predict efficiently the incoming content
requests and adapts strongly with the content requests
while others neglect or could not adapt with the
temporal locality of content. CafRepCache is
followed by LocationCache and SocialCache that
manage up to 76.6% cache hit ratio regarding the
increase of content popularity skewness. We observe
that LTRU and OCPCP achieve the lowest cache hit
ratio (ranging from 24,3% to 60%) as it relies only on
simple request frequency or recency metric, thus
could not be able to predict adaptively the temporal
locality of content request patterns.
As the content popularity changes in real-time
regarding the variation in user interests, we further
study the impact of varying content popularity
fluctuation on the caching performance in order to
evaluate the effectiveness of the content request
recency heuristic. The Fig.7 shows the popularity
variation range from 20 to 120 popularity rank for
each round of the experiment. As caching
performance of slight variation on content popularity
has advantage over sharp variation, it is
Interdependent Multi-layer Spatial Temporal-based Caching in Heterogeneous Mobile Edge and Fog Networks
41
Figure 6: Cache hit ratio vs Popularity zipf alpha.
challenging for the caching algorithms to adapt
quickly to the severe and quick popularity alteration.
As shown in Fig.7, CafRepCache’s request
recency metric keeps good stability regarding the
popularity change (typically above 83.1% cache hit
ratio), predict adaptively the emerging contents that
may become highly popular in near future and
contents that are currently considered as high popular
but will be less popular soon. The sensitiveness of our
proposed algorithm provides better adaptation to
rapidly changing content popularity and different
network environment while all other competitive
caching algorithms could not be able to adapt with
sharp variation in popularity.
Figure 7: Cache hit ratio vs Popularity variation.
We vary the cache size and measure the cache
eviction ratio to evaluate the effectiveness of the
content request betweeness heuristic. Eviction is one
of the popular metrics for the performance evaluation
of content caching. When the cache space is full and
caching point makes a decision to cache a new arrived
content then one of the cached contents is evicted or
offloaded to free up the buffer. When the resource is
limited and the eviction rate is potentially high due to
one-timer contents, the overall network throughput is
affected in terms of cache hit ratio and content
retrieval latency. In other words, if a stored content is
evicted incorrectly and a new request arrives for it,
then there is a cache miss and the content has to be
retrieved from other nodes or publishers directly. As
a result, this increases the content retrieval latency
and decreases the cache hit ratio of the caching
services. Smaller cache buffer size offers more
selective cached contents, thus requires more accurate
content popularity prediction.
Fig. 8 shows that CafRepCache achieves the
lowest eviction ratio regarding the dynamic changing
in size of the caching points (decreasing from 0.62 to
0.34 eviction ratio when the cache size is increased)
compared to the state-of-the-arts caching algorithms.
CafRepCache is followed by SocialCache and
LocationCache (ranging from 0.75 to 0.46 eviction
ratio) while TLRU and OCPCP have the worst
performance, especially when the cache space is
relatively small. This is due to the request betweeness
heuristic allows CafRepCache balance the trade-off
between current observed content popularity versus
long terms interest in order to avoid caching quickly
potentially one-timer contents or fake news and
losing the long-term useful contents.
Figure 8: Eviction rate vs Relative cache size.
In order to understand the request spatial-based
heuristics integrated in our caching algorithm, we
vary the content request patterns which follows
PECCS 2019 - 9th International Conference on Pervasive and Embedded Computing and Communication Systems
42
different spatial localisation factor β (0.1-0.85)
utilized in Hawkess process (Dabirmoghaddam et al.,
2014) and measure the cache hit ratio. In line with
(Dabirmoghaddam et al., 2014), with low localization
factor β, content requests are generated
independently. The generated trace, therefore,
conforms to the IRM assumption and hence, one
single time content is considered. With a localization
factor of 0.85, other near nodes in a region more likely
request the same content after a node requests it.
Fig. 9 shows that the request spatial clustering
heuristic enables CafRepCache to adapt with the
spatial locality of content requests. CafRepCache
achieves the best performance (around 87% cache hit
ratio) followed by LocationCache (80%) and
SocialCache (72%). TLRU and OCPCP have no (or
little) improvement to the accuracy of predicting
content popularity, ranging from 52% to 63% cache
hit ratio regarding the dynamically changing of
spatial localization factor.
We observe that higher β even leads to bigger gap
between CafRepCache and other state-of-the-art
content solutions. It is due to the spatial locality
heuristic helps to classify and recognise content
interests coming from localised group of subscribers,
then adaptively predict that the contents will be
requested again by other subscribers within that
location.
Figure 9: Cache hit ratio vs Spatial localisation factor.
We evaluate the effect of subscriber-caching point
connectivity on the performance of different state-of-
the-art caching protocols in order to understand the
nature of caching points regarding the dynamic
connectivity of mobile subscribers.
Fig. 10 shows that CafRepCache outperforms
other competitive caching protocols in terms of
average delay measured by the number of hops.
Figure 10: Average latency (hops) vs. Subscriber-Caching
point connectivity.
CafRepCache is able to bring the cached contents to
only a few hops away from subscribers (2.7 hops)
regarding the dynamic centrality of caching points.
This is due to CafRepCache benefits from
multidimensional multilayer analytics that allow it to
place the most suitable set of contents in the most
suitable set of caching points which are not only
highly central but also have similarity in contacts and
interest requests with the subscribers.
5 CONCLUSIONS
We proposed multilayer adaptive predictive
distributed collaborative analytics and heuristics for
enabling spatial-temporal locality awareness of
mobility and traffic patterns as well as mobility-
content traffic interplay for opportunistic caching in
mobile edge/fog networks. Our combined heuristics
allow the caching protocol to be more flexible and
responsive to the dynamic mobility and complex
content request patterns in the unreliable scenarios
imposed by varying publishers and subscribers as
well as dynamic resource availability. We performed
extensive real trace-driven experiments in ONE
simulation (Keränen et al., 2009) and showed that the
proposed predictive heuristics help CafRepCache
caching framework to perform better than the state-
of-the-art caching solutions in terms of success ratio,
cache hit ratio, average delay and eviction rate.
We aim to explore our ego-network analytics
and heuristics in greater detail and propose adaptive
context-aware weighting of the complementary
analytics and utilities. We plan to deploy our
approach in different application scenarios which
Interdependent Multi-layer Spatial Temporal-based Caching in Heterogeneous Mobile Edge and Fog Networks
43
have complex temporal network topologies such as
smart agriculture (Wietrzyk and Radenkovic, 2010;
Brun-Laguna et al., 2016), urban emergency (Huynh
and Radenkovic, 2017), intelligent transport system
(Loscri et al., 2019), and smart manufacturing
(Radenkovic et al., 2015) using software-defined
networking (SDN) or network function virtualization
(NFV) as in (Radenkovic, 2016; Radenkovic and
Huynh, 2016).
REFERENCES
Bilal, M. and Kang, S.-G. (2014) Time aware least recent
used (TLRU) cache management policy in ICN, 16th
International Conference on Advanced Communication
Technology. IEEE.
Bornholdt, S. and Schuster, H. G. (2006) Handbook of
graphs and networks: from the genome to the internet.
John Wiley & Sons.
Breslau, L., Cao, P., Fan, L., Phillips, G. and Shenker, S.
(1999) Web caching and Zipf-like distributions:
Evidence and implications, Ieee Infocom. INSTITUTE
OF ELECTRICAL ENGINEERS INC (IEEE).
Brun-Laguna K., Diedrichs A. L., Dujovne D., Leone R.,
Vilajosana X. and Watteyne T. (2016). (Not so)
intuitive results from a smart agriculture low-power
wireless mesh deployment. Proceedings of the Eleventh
ACM Workshop on Challenged Networks. ACM.
Dabirmoghaddam, A., Barijough, M. M. and Garcia-Luna-
Aceves, J. (2014) Understanding optimal caching and
opportunistic caching at the edge of information-centric
networks, Proceedings of the 1st ACM conference on
information-centric networking. ACM.
Daly, E. M. and Haahr, M. (2007) Social network analysis
for routing in disconnected delay-tolerant manets,
Proceedings of the 8th ACM international symposium
on Mobile ad hoc networking and computing. ACM.
Dán, G. and Carlsson, N. (2010) Power-law revisited: large
scale measurement study of P2P content popularity,
IPTPS.
D’Silva, K., Noulas, A., Musolesi, M., Mascolo, C. and
Sklar, M. (2018) Predicting the temporal activity
patterns of new venues. EPJ Data Science, 7(1), 13.
Flores, H., Sharma, R., Ferreira, D., Kostakos, V., Manner,
J., Tarkoma, S., Hui, P. and Li, Y. (2017) Social-aware
hybrid mobile offloading. Pervasive and Mobile
Computing, 36, 25-43.
Fricker, C., Robert, P. and Roberts, J. (2012) A versatile
and accurate approximation for LRU cache
performance, 2012 24th International Teletraffic
Congress (ITC 24). IEEE.
Huynh, V. S. H. and Radenkovic M. (2017). A novel cross-
layer framework for large scale emergency
communications. 2017 13th International Wireless
Communications and Mobile Computing Conference
(IWCMC), IEEE.
Huynh, V. S. H. and Radenkovic, M. (2018).
Understanding information centric layer of adaptive
collaborative caching framework in mobile
disconnection-prone networks. 2018 14th International
Wireless Communications & Mobile Computing
Conference (IWCMC), IEEE
Keränen, A., Ott, J. and Kärkkäinen, T. (2009) The ONE
simulator for DTN protocol evaluation, Proceedings of
the 2nd international conference on simulation tools
and techniques. ICST (Institute for Computer Sciences,
Social-Informatics and ….
Le, T., Lu, Y. and Gerla, M. (2015) Social caching and
content retrieval in disruption tolerant networks
(DTNs), 2015 International Conference on Computing,
Networking and Communications (ICNC). IEEE.
Liu, Z., Yang, X., Yang, Y., Wang, K. and Mao, G. (2018)
DATS: Dispersive Stable Task Scheduling in
Heterogeneous Fog Networks. IEEE Internet of Things
Journal.
Loscri, V., Manzoni P., Nitti M., Ruggeri G. and Vegni A.
(2019). A social internet of vehicles sharing SIoT
relationships.
Mardham, D., Madria, S., Milligan, J. and Linderman, M.
(2018) Opportunistic distributed caching for mission-
oriented delay-tolerant networks, 2018 14th Annual
Conference on Wireless On-demand Network Systems
and Services (WONS). IEEE.
Nicosia, V., Tang, J., Mascolo, C., Musolesi, M., Russo, G.
and Latora, V. (2013) Graph metrics for temporal
networks, Temporal networksSpringer, 15-40.
Radenkovic, M. (2016). Cognitive privacy for personal
clouds. Mobile Information Systems 2016.
Radenkovic, M. and Grundy, A. (2011) Congestion aware
forwarding in delay tolerant and social opportunistic
networks, 2011 Eighth International Conference on
Wireless On-Demand Network Systems and Services.
IEEE.
Radenkovic, M. and Huynh V. S. H. (2016). Low-cost
mobile personal clouds. 2016 International Wireless
Communications and Mobile Computing Conference
(IWCMC), IEEE.
Radenkovic, M. and Huynh, V. S. H. (2017) Collaborative
cognitive content dissemination and query in
heterogeneous mobile opportunistic networks,
Proceedings of the 3rd Workshop on Experiences with
the Design and Implementation of Smart Objects.
ACM.
Radenkovic, M., Huynh, V. S. H. and Manzoni, P. (2018)
Adaptive real-time predictive collaborative content
discovery and retrieval in mobile disconnection prone
networks. IEEE Access, 6, 32188-32206.
Radenkovic, M., Kostadinov I. and Wietrzyk B. (2015).
Increasing communication reliability in manufacturing
environments. 2015 International Wireless
Communications and Mobile Computing Conference
(IWCMC), IEEE.
Ruan, M., Chen, X. and Zhou, H. (2019) Centrality
prediction based on K-order Markov chain in Mobile
Social Networks. Peer-to-Peer Networking and
Applications, 1-11.
PECCS 2019 - 9th International Conference on Pervasive and Embedded Computing and Communication Systems
44
Saha, S., Lukyanenko, A. and Ylä-Jääski, A. (2013)
Cooperative caching through routing control in
information-centric networks, 2013 Proceedings IEEE
INFOCOM. IEEE.
Scott, J., Gass, R., Crowcroft, J., Hui, P., Diot, C. and
Chaintreau, A. (2006) CRAWDAD dataset
cambridge/haggle (v. 2006-09-15). CRAWDAD
wireless network data archive.
Wang, L., Tyson, G., Kangasharju, J., Crowcroft, J., Wang,
L., Tyson, G., Kangasharju, J. and Crowcroft, J. (2017)
Milking the cache cow with fairness in mind.
IEEE/ACM Transactions on Networking (TON), 25(5),
2686-2700.
Wang, T., Hui, P., Kulkarni, S. and Cuff, P. (2014)
Cooperative caching based on file popularity ranking in
delay tolerant networks. arXiv preprint
arXiv:1409.7047.
Yang, D., Zhang, D., Zheng, V. W. and Yu, Z. (2014)
Modeling user activity preference by leveraging user
spatial temporal characteristics in LBSNs. IEEE
Transactions on Systems, Man, and Cybernetics:
Systems, 45(1), 129-142.
Yoneki, E., Hui, P. and Crowcroft, J. (2008) Distinct types
of hubs in human dynamic networks, Proceedings of
the 1st Workshop on social Network Systems. ACM.
Zhang, G., Tang, B., Wang, X. and Wu, Y. (2014) An
optimal cache placement strategy based on content
popularity in content centric network. JOURNAL OF
INFORMATION & COMPUTATIONAL SCIENCE,
11(8), 2759-2769.
Wietrzyk, B. and M. Radenkovic (2010). Realistic large
scale ad hoc animal monitoring. International Journal
On Advances in Life Sciences 2(1 and 2).
Interdependent Multi-layer Spatial Temporal-based Caching in Heterogeneous Mobile Edge and Fog Networks
45