Caching Meets Millimeter Wave Communications For

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2771419, IEEE
Transactions on Wireless Communications
Caching Meets Millimeter Wave Communications for

Enhanced Mobility Management in 5G Networks
Omid Semiari1 , Walid Saad2 , Mehdi Bennis3 , and Behrouz Maham4
1
Department of Electrical Engineering, Georgia Southern University, Statesboro, GA, USA, Email: osemiari@georgiasouthern.edu
2
Department of Electrical and Computer Engineering, Virginia Tech, Blacksburg, VA, USA, Email: walids@vt.edu
3
Centre for Wireless Communications, University of Oulu, Finland, Email: bennis@ee.oulu.fi
4
Department of Electrical and Electronic Engineering, Nazarbayev University, Astana, Kazakhstan, Email: behrouz.maham@nu.edu.kz
Abstract—One of the most promising approaches to overcome wireless cellular networks. To cope with this traffic increase,
the uncertainty and dynamic channel variations of millimeter several new technologies are anticipated for 5G cellular sys-
wave (mmWave) communications is to deploy dual-mode base tems: 1) dense deployment of small cell base stations (SBSs),
stations that integrate both mmWave and microwave (µW)
frequencies. If properly designed, such dual-mode base stations 2) exploitation of the large amount of available bandwidth at
can enhance mobility and handover in highly mobile wireless millimeter wave (mmWave) frequencies, and 3) edge content
environments. In this paper, a novel approach for analyzing caching directly at the user equipments (UEs) to reduce delay
and managing mobility in joint mmWave-µW networks is pro- and improve quality-of-service (QoS). The dense deployment
posed. The proposed approach leverages device-level caching of SBSs with reduced cell-sizes will boost the capacity of
along with the capabilities of dual-mode small base stations
(SBSs) to minimize handover failures, reduce inter-frequency wireless networks by decreasing UE-SBS distance, removing
measurement energy consumption, and provide seamless mobility coverage holes, and improving spectral efficiency. Meanwhile,
in emerging dense heterogeneous networks. First, fundamental mmWave communications will provide high data rates by
results on the caching capabilities, including caching probability leveraging directional antennas and enabling transmission over
and cache duration are derived for the proposed dual-mode a large bandwidth that can reach up to 5 GHz. In addition,
network scenario. Second, the average achievable rate of caching
is derived for mobile users. Moreover, the impact of caching on exploiting the high storage capacity of modern smartphones
the number of handovers (HOs), energy consumption, and the to cache the data at the UE increases the flexibility and
average handover failure (HOF) is analyzed. Then, the proposed robustness of resource management, in particular, for mobile
cache-enabled mobility management problem is formulated as a UEs (MUEs). In fact, caching allows the network to store
dynamic matching game between mobile user equipments (MUEs) the data content in advance, while enabling MUEs to use
and SBSs. The goal of this game is to find a distributed HO
mechanism that, under network constraints on HOFs and limited the cached content when sufficient wireless resources are not
cache sizes, allows each MUE to choose between: a) executing available.
an HO to a target SBS, b) being connected to the macrocell However, dense heterogeneous networks (HetNets), com-
base station (MBS), or c) perform a transparent HO by using posed of macrocell base stations (MBSs) and SBSs with
the cached content. The formulated matching game inherently various cell sizes, will introduce three practical challenges for
captures the dynamics of the mobility management problem
caused by HOFs. To solve this dynamic matching problem, a mobility management. First, MUEs will experience frequent
novel algorithm is proposed and its convergence to a two-sided handovers (HOs), while passing SBSs with relatively small
dynamically stable HO policy for MUEs and target SBSs is cell sizes, which naturally increases the overhead and delay
proved. Numerical results corroborate the analytical derivations in HetNets. Such frequent HOs will also increase handover
and show that the proposed solution will significantly reduce failure (HOF), particularly for MUEs that are moving at high
both the HOF and energy consumption of MUEs, resulting in
an enhanced mobility management for heterogeneous wireless speeds [1]. In fact, due to the small and disparate cell sizes in
networks with mmWave capabilities. HetNets, MUEs will not be able to successfully finish the HO
Index Terms— millimeter wave communications; game the- process by the time they trigger HO and pass a target SBS.
ory; resource allocation. Second, the inter-frequency measurements that are needed to
discover target SBSs can be excessively power consuming
I. I NTRODUCTION and detrimental for the battery life of MUEs, especially in
dense HetNets with frequent HOs. Third, microwave (µW)
The proliferation of bandwidth-intensive wireless applica-
frequencies are stringently congested, and thus, frequent HOs
tions such as social networking, high definition video stream-
may introduce unacceptable overhead and limit the available
ing, and mobile TV has drastically strained the capacity of
frequency resources for the static users. In this regard, offload-
This research was supported by the U.S. National Science Foundation under ing MUEs from heavily used µW frequencies to mmWave
Grants CNS-1460316, CNS-1513697, and IIS-1633363. frequencies can substantially improve the spectral efficiency
1536-1276 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2771419, IEEE
at the µW network. ongoing IP service can be cached in advance and continuously

To address these challenges and enhance mobility manage- transferred among different data centers as MUEs move across
ment in HetNets, an extensive body of work has appeared different cells. In [16], the authors propose a caching frame-
in the literature [2]–[17]. In [2], the authors provide a com- work that stores different parts of a content at different base
prehensive overview on mobility management in IP networks. stations, allowing MUEs to randomly move across different
The authors in [3] present different distributed mobility man- cells and download different cached parts of the original
agement protocols at the transport layer for future dense content whenever possible. In addition, in [17], a proactive
HetNets. In [4], an energy-efficient SBS discovery method caching solution is proposed for mobility management by
is proposed for HetNets. The work in [5] investigates HO exploiting MUEs’ trajectory information. Despite addressing
decision algorithms that focus on improving HO between important problems, the body of work in [14]–[17] relies on
femtocells and LTE-Advanced systems. In [6], the authors protocols designed for higher network layers. Moreover, these
overview existing approaches for vertical handover decisions solutions do not consider caching directly at the MUEs and
in HetNets. In [7], the authors study the impact of channel focus on mobility management at the core network. However,
fading on mobility management in HetNets. In addition, the we will show how leveraging high capacity mmWave com-
work in [7] shows that increasing the sampling period for munication can complement the notion of caching at MUEs.
HO decision decreases the fading impact, while increasing the In addition, caching at MUEs will provide opportunities to
ping-pong effect. In [8], the authors propose an HO scheme perform transparent HOs in HetNets, without requiring any
that takes into account the speed of MUEs to decrease frequent data session with a target SBS.
HOs in HetNets. The authors in [9] propose an HO scheme The main contribution of this paper is a novel mobility
that supports soft HO by allowing MUEs to connect with management framework that addresses critical handover is-
both a macrocell base station (MBS) and SBSs. Furthermore, sues, including frequent HOs, HOF, and excessive energy
a distributed mobility management framework is proposed in consumption for seamless HO in emerging dense cellular
[10] which uses multiple frequency bands to decouple the data networks with mmWave capabilities. In fact, we propose a
and control planes. model that allows MUEs to cache their requested content
Although interesting, the body of work in [2]–[10] does not by exploiting high capacity mmWave connectivity whenever
consider mmWave communications and caching capabilities available. As such, the MUEs will use the cached content
for mobility management and solely focuses on HetNets oper- and avoid performing any HO, while passing SBSs with
ating over µW frequencies. In addition, it does not study the relatively small cell sizes. First, we propose a geometric
opportunities that caching techniques can provide for mobility model to derive tractable, closed-form expressions for key
management. In [11], an HO scheme for mmWave networks performance metrics, including the probability of caching,
is proposed in which the MBS acts as an anchor for mmWave cumulative distribution function of caching duration, and the
SBSs to manage control signals. However, [11] assumes that average data rate for caching at an MUE over a mmWave
line-of-sight (LoS) mmWave links are always available and link. Moreover, we shed light on how to leverage caching
provides no analytical results to capture the directional nature in mmWave-µW networks in order to reduce the number of
of mmWave communications. In [12], the authors propose HOs and the average HOF. Then, we formulate the proposed
a resource allocation scheme for hybrid mmWave-µW net- cache-enabled mobility management framework as a dynamic
works that enhances video streaming by buffering content matching game, so as to provide a distributed solution for
over mmWave links. However, [12] does not address any mobility management in HetNets, while taking the dynamics
mobility management challenge, such as frequent HOs or of the system into account. To solve the formulated dynamic
HOF. Our early work in [13] provided some of the basic matching problem, we first show that conventional algorithms
insights on mobility management in µW-mmWave networks. such as the deferred acceptance algorithm adopted in [20] and
However, in contrast to this work, [13] solely focuses on an [21], fail to guarantee a dynamically stable HO between MUEs
average performance analysis, does not consider dynamic HO and SBSs. Therefore, we propose a novel distributed algorithm
problem for multi-MUE scenarios, and does not propose any that is guaranteed to converge to a dynamically stable HO
energy management mechanisms for handling inter-frequency policy in dense HetNets. Subsequently, the complexity of the
measurements. proposed algorithm in terms of signaling overhead is analyzed.
Proactive caching for enhancing mobility management has Under practical settings, we show that the proposed cache-
been motivated by the works in [14]–[19]. In [14], the authors enabled HO framework can decrease the average HOF rate by
discuss the potential of content caching at either evolved up to 45%, even for MUEs with high speeds. In addition,
packet core network or radio access network to minimize the simulation results provide insights on the achievable gains
traffic overhead at the core network. Moreover, the authors by the proposed distributed algorithm, in terms of reducing
in [15] propose a proactive caching framework in which an energy consumption for cell search, as well as improving
traffic offload from the µW frequencies.

The rest of this paper is organized as follows. Section II
presents the system model. Section III presents the analysis for
caching in mobility management. Performance analysis of the
cache-enabled mobility management is provided in Section IV.
Section V formulates the mobility management as a dynamic
matching and presents the proposed algorithm. Simulation
results are presented in Section VI and conclusions are drawn
in Section VII.
II. S YSTEM M ODEL
Consider a HetNet composed of an MBS and K SBSs
within a set K distributed uniformly across an area. Each SBS
k ∈ K can be viewed as a picocell or a femtocell, depending
on its transmit power pk . Picocells are typically deployed in
outdoor venues while femtocells are relatively low-power and
suitable for indoor deployments. We consider dedicated carrier
deployments [4], [22]–[24], in which SBSs operate at µW Fig. 1: SBSs coverage with RSS threshold of −80 dB. Red circles
frequencies that are different than those used by the MBS and, show the simplified cell boundaries.
thus, there is no interference between SBSs and the MBS 1 .
The SBSs are also equipped with mmWave front-ends to serve
where (1) holds for ru,k ≥ rref , with rref and ru,k denoting,
MUEs over either mmWave or µW frequency bands [25]. The
respectively, the reference distance and distance between the
dual-mode capability allows to integrate mmWave and µW
MUE u and SBS k. In addition, α is the path loss exponent,
radio access technologies (RATs) at the medium access control
λ is the wavelength at carrier frequency fc = 73 GHz over
(MAC) layer of the radio interface and reduce the delay and
the E-band, and χ is a Gaussian random variable with zero
overhead for fast vertical handovers between both RATs [25].
mean and variance ξ 2 . The path loss parameters α and ξ
Thus, MUEs can receive their requested traffic over either the
will have different values, depending on whether the mmWave
mmWave or the µW band. Within this network, we consider
link is line-of-sight (LoS) or non-LoS (NLoS). Over the µW
a random direction (RD) mobility model [26] in which a set
frequency band, the path loss model follows (1), however, with
U of U MUEs are distributed randomly and move across a
parameters that are specific to sub-6 GHz frequencies.
geographical area during a time frame T . In fact, each MUE
u ∈ U moves with an average speed vu ∈ [vmin , vmax ] in a An illustration of the considered HetNet is shown in Fig.
random direction θu ∈ [0, 2π], with respect to the θ = 0 1. The coverage for each SBS at the µW frequency is shown
horizontal angle. It is assumed that vu does not change during based on the maximum received signal strength (max-RSS)
the considered time frame T . This model does not make any criteria with a threshold of −80 dB. White spaces in Fig. 1
restrictive assumptions on the location of obstacles (such as delineate the areas that are covered solely by the MBS. Here,
buildings) or the road pattern, and thus, it can capture a general we observe that shadowing and small-scale fading effects can
case in which it is equally probable for the MUE to move in adversely increase the ping-pong effect for MUEs. To cope
different directions. with this issue, the 3GPP standard suggests L1/L3 filtering
which basically applies averaging to the RSS samples, in both
A. Channel model time and frequency, as explained in [1].
The large-scale channel effect over mmWave frequencies
for a link between an SBS k and an MUE u ∈ U, in dB, is B. Antenna model and configuration
given by2 : To overcome the excessive path loss at the mmWave fre-

4πr0 ru,k quency band, the MUEs will be equipped with electronically
L(u, k) = 20 log10 +10α log10 +χ, (1)
λ r0 steerable antennas which allow them to achieve beamform-
ing gains at a desired direction. The antenna gain pattern
1 The key reason for considering different frequency bands for the MBS and
for MUEs follows the simple and widely-adopted sectorized
SBSs is that minimizing the excessive power to perform inter-frequency small
cell measurements is one of the main objectives of 3GPP Evolved Universal
pattern which is given by [28]–[34]:
Terrestrial Radio Access (E-UTRA), as reported in [24]. (
2 The free space path loss model in (1) has been adopted in many existing Gmax , if θ < |θm |,
works, such as in [27], that carry out real-world measurements to characterize G(θ) = (2)
mmWave large scale channel effects.
Gmin , otherwise,
TABLE I: Variables and notations

Notation Description Notation Description
K Number of SBSs K Set of SBSs
U Number of MUEs U Set of MUEs
θu Moving angle of MUEs vu Speed of MUEs
pk Transmit power of SBS k B Segment size of video (bits)
Ωu Cache size of MUE u Ωumax Maximum cache size
tcu Caching duration of MUE u Q Video play rate
c
R̄ (u, k) Average achievable caching rate dc Traversed distance using cached content
∆T Time-to-trigger (TTT) rc Traversed distance in caching duration
Ts Inter-frequency cell scanning interval tMTS Minimum time-of-stay (ToS)
θk Beamwidth for SBS k Es Consumed energy per cell search
tu,k Time-of-stay for MUE u at SBS k tMTS Minimum required time-of-stay
where θ and θm denote, respectively, the azimuth angle and addition, tcu is the caching duration which is equal to the time
the antennas’ main lobe beamwidth. Gmax and Gmin denote, needed for a given MUE u to traverse the mmWave beam of its
respectively, the antenna gain of the main lobe and side lobes. serving SBS. To derive (3), we note that, given the achievable
For SBSs, we use a model similar to the sectorized pattern rate R̄c (u, k) and the time tcu , the amount of data (bits) that
in (2), however, we allow each SBS k to form Nk beams, can be cached at a given MUE is R̄c (u, k)tcu . Moreover, since
either by using Nk antenna arrays or forming multi-beam each video segment has a size ofB bits, the total number of
beamforming. The beam patten configuration of an SBS k segments that can be cached is
R̄c (u,k)tcu
. Given that any
B
is shown in Fig. 2, where Nk = 3 equidistant beams in
θ ∈ [0, 2π] are formed. To avoid the complexity and overhead device can store a limited number of video segments due to
of beam-tracking for mobile users, the direction of the SBSs’ its limited maximum cache size, Ωumax , the overall number of
beams in azimuth is fixed3 . In fact, an MUE can connect to cached segments will be given by (3). Considering the small
an SBS k over a mmWave link, if the MUE traverses the area green triangle in Fig. 2 as the current location of a given
covered by the k’s mmWave beams. It is assumed that for a MUE crossing a mmWave beam, the caching duration will
desired link between an SBS k and an MUE u, the overall be tcu = ruc /vu where ruc is the distance traversed across the
transmit-receive gain is ψu,k = G2max . mmWave beam. Moreover, R̄c (u, k) is the average achievable
rate for the MUE u during tcu . Given Ωu (k) and the video play
C. Traffic model rate of Q, specified for each video content, the distance an
Video streaming is one of the wireless services with most MUE u can traverse with speed vu , while playing the cached
stringent quality-of-service (QoS) requirement. Meeting the video content will be
QoS demands of such services is prone to the delay caused Ωu (k)
by frequent handovers in HetNets. In addition, HOFs can dc (u, k) = vu . (4)
Q
significantly degrade the performance by making frequent ser-
vice interruptions. Therefore, our goal is to enhance mobility In fact, the MUE can traverse a distance dc (u, k) by using the
management for MUEs that request video or streaming traffic. cached video content after leaving its serving cell k, without
Each video content is partitioned into small segments, each of requiring an HO to any of the target cells. Meanwhile, the
size B bits [35]–[37]. The network incorporates caching to location information and control signals, such as paging, can
transmit incoming video segments to an MUE, whenever a be handled by the MBS during this time. As we discuss in
high capacity mmWave connection is available. In fact, high details, such caching mechanism will help MUEs to avoid re-
capacity mmWave connection, if available, allows to cache a dundant cell search and HOs, resulting in an efficient mobility
large portion or even the entire video in a very short period management in dense HetNets.
of time. We define the cache size of Ωu (k) for an arbitrary
MUE u, associated with an SBS k, as the number of video D. Handover procedure and performance metrics
segments that can be cached at MUE u as follows: The HO process in the 3GPP standard proceeds as follows:
n R̄c (u, k)tc
u
o 1) Each MUE will do a cell search every Ts seconds, which
Ωu (k) = min , Ωumax , (3) can be configured by the network or directly by the MUEs, 2)
B
If any target cell offers an RSS plus a hysteresis that is higher
where b.c and min{., .} denote, respectively, the floor and than the serving cell, even after L1/L3 filtering of input RSS
minimum operands and Ωumax is the maximum cache size. In samples, the MUE will wait for a time-to-trigger (TTT) of ∆T
3 Nonetheless, the model can be extended to accommodate beamforming seconds to measure the average RSS from the target cell, 3) If
with dynamic parameters. the average RSS is higher than that of the serving SBS during
s
Note that increasing Ts reduces Etotal , which is desirable.
However, less frequent scans will be equivalent to less HOs
to SBSs. Therefore, there is a tradeoff between reducing
the consumed power for cell search and maximizing traffic
offloads from the MBS to SBSs. Content caching will allow
increasing Ts , while maintaining traffic offloads from the
MBS.
Next, we propose a geometric framework to analyze the
caching opportunities, in terms of the caching duration tc , and
the average achievable rate R̄c , for MUEs moving at random
directions in joint mmWave-µW HetNets.
III. A NALYSIS OF M OBILITY M ANAGEMENT WITH

C ACHING C APABILITIES
Fig. 2: Antenna beam configuration of a dual-mode SBS with Nk = In this section, we first investigate the probability of serving
3. Shaded areas show the mmWave beams.
an arbitrary MUE over mmWave frequencies by a dual-mode
SBS.
TTT, the MUE triggers HO and sends the measurement report
A. Probability of mmWave coverage
to its serving cell. The averaging over the TTT duration will
reduce the ping-pong effect resulting from instantaneous CSI In Fig. 2, the small circle represents the intersection of an
variations, and 4) HO will be executed after the serving SBS arbitrary MUE’s trajectory with the coverage area of an SBS
sends the HO information to the target SBS. k. In this regard, Pck (Nk , θk ) represents the probability that the
In our model, we modify the above HO procedure to lever- MUE with a random direction θ crosses the mmWave coverage
age the caching capabilities of MUEs during mobility. Here, areas of SBS k. From Fig. 2, we observe that the MUE will
we let each MUE u dynamically determine Ts , depending on pass through the area within mmWave coverage only if this
the cache size Ωu , the video play rate Q, and the MUE’s speed MUE’s direction is inside the angle AoB. [ Hence, we can state
vu . That is, an MUE u is capable of muting the cell search the following.
while Ωu /Q is greater than ∆T , which enables it to have ∆T
Theorem 1. If an SBS k has formed a mmWave beam pattern
seconds to search for a target SBS before the cached content
with Nk ≥ 2 main lobes, each with a beamwidth θk > 0, the
runs out.
probability of content caching will be given by:
Next, we consider the HOF as one of the key performance
Pk (Nk , θk ) = 2π + 1 − 2π 2 1 − N + 4π .

metrics for any HO procedure. One of the main reasons for the c Nk θk Nk θk 1 1 θk
potential increase in HOF in HetNets is due to the relatively k
small cell sizes, compared to MBS coverage. In fact, HOF (7)
is typical if the time-of-stay (ToS) for an MUE is less than
Proof. See Appendix A.
the minimum ToS (MTS) required for performing a successful
HO. That is, We can verify (7) by considering an example scenario with
Nk = 3 and θk = 2π 3 . For this example, (7) results in
(
γHOF (u, k) =
1, if tu,k < tMTS ,
0, otherwise,
(5) Pck (Nk , θk ) = 1 which correctly captures the fact that the
entire cell is covered by mmWave beams.
where tu,k is the ToS for MUE u to pass across SBS k
coverage. Although a short ToS may not be the only cause B. Cumulative distribution function of the caching duration
for HOFs, it becomes very critical within an ultra dense small To enable an MUE to use the cached content while not being
cell network that encompasses MUEs moving at high speeds associated to an SBS, it is critical to analyze the distribution
[24]. of caching duration tc for an arbitrary MUE with a random
To search the µW carrier for synchronization signals and direction and speed. In this regard, consider the small green
decode the broadcast channel (system information) of the triangle in Fig. 2, which represents the location of an arbitrary
detected SBSs, the MUEs have to spend an energy E s per MUE u, xu = (xu , yu ) ∈ R2 , crossing a mmWave beam.
each cell search [4]. Hence, the total energy consumed by an First, we note that the geometry of the mmWave beam of any
MUE for cell search during time T will be given SBS can be defined by the location of the SBS, as well
T as the sides of the beam angle. Without loss of generality, we
s
Etotal = Es . (6) assume that the SBS of interest is located at the center, such
Ts
that xk = (0, 0). Therefore, the two sides of the beam angle
1
will be given by
0.9
y = x tan(θ0 − θk ), y = x tan(θ0 ), x > 0. (8) 0.8
Assuming that the MUE u is currently located on the angle 0.7
side x = y cos(θ0 −θk ), as shown by the

small triangle
in Fig. 0.6
xu
2, then θ0 in (8) will be θ0 = arccos ru,k (xu ) + θk , where
CDF
0.5
p
ru,k (x) = x2u + yu2 . Hereinafter, we will use the parameter 0.4
ru,k = 10
θ0 to simplify our analysis. Let Ftc (.) be the cumulative 0.3 ru,k = 50
distribution function (CDF) of the caching duration tc . Thus, ru,k = 100
0.2
ru,k = 150
Ftcu (t0 ) = P(tcu ≤ t0 ) = P(ruc ≤ vu t0 ), (9) 0.1 ru,k = 200
where ruc is the distance that MUE u will traverse across the 0
0 2 4 6 8 10 12
Caching duration (s)
mmWave beam, as shown in Fig. 2. Given the location of
MUE xu , the minimum possible distance to traverse, rumin , is Fig. 3: CDF of caching duration tc .

xu tan θ0 − yu
min
ru = p . (10) IV. P ERFORMANCE A NALYSIS OF THE P ROPOSED
1 + tan2 θ0 C ACHE - ENABLED M OBILITY M ANAGEMENT S CHEME
In fact, (10) represents the distance of the point xu from Next, we analyze the average achievable rate for content
the beam angle side y = x tan(θ0 ). If rumin > vu t0 , then caching, for an MUE with speed vu , direction θu , and initial
Ftcu (t0 ) = 0. Therefore, for the remainder of this analysis we distance ru,k (x) from the serving dual-mode SBS. In addition,
consider rumin ≤ vu t0 . Next, let x0u denote the intersection we evaluate the impact of caching on mobility management.
of the MUE’s path with line y = x tan(θ0 ). It is easy For this analysis, we ignore the shadowing effect and only
to see that x0u = (xu + ruc cos θu , yu + ruc sin θu ). Hence, consider distance path loss.
yu + ruc sin θu = [xu + ruc cos θu ] tan θ0 , and ruc , i.e., the A. Average Achievable Rate for Caching
distance that MUE u traverses during the caching duration
tc , is given by: The achievable rate of caching is given by:
Z ru,k (x0 ) −α
!
yu − xu tan θ0 c 1 βPt ψru,k
ruc = vu tcu = . (11) R (u, k) = w log 1 + dru,k , (14)
tan θ0 cos θu − sin θu vu tcu ru,k (x) wN0
λ
Next, from (9) and (11), the CDF can be written as where β = ( 4πr 0
)2 r0α . The integral in (14) is taken over the
line with length ruc that connects the MUE location x to x0 ,
yu − xu tan θ0
Ftcu (t0 ) = P ≤ vu t0 . (12) as shown in Fig. 2. With this in mind, we can find the average
tan θ0 cos θu − sin θu achievable rate of caching R̄c as follows.
Using the geometry shown in Fig. 2, we find the CDF of the Theorem 2. The average achievable rate for an MUE u served
caching duration as follows: by an SBS k, R̄c (u, k), is:
Lemma 1. The CDF of the caching duration, tc , for an R̄c (u, k) = Pck (Nk , θk )Rc (u, k), (15)
arbitrary MUE u with speed vu is given by Z f (0)
1
1
min
ru = δ2 2 (θ)
log (1 + δ1 f α (θ)) df (θ), (16)
Ftc (t0 ) = arccos + (13) f (θk ) f
π − θk vu t0 p
ln(δ1 f 2 (θk ) + 1)
(a) δ2 p
min
ru
min
ru = 2 δ1 arctan( δ1 f (θk ))− ,
min arccos , arccos . ln(2) f (θk )
ru,k (x) vu t0 p p ln(δ1 f 2 (0) + 1)

−2 δ1 arctan( δ1 f (0))+ , (17)
Proof. See Appendix B. f (0)
h i−α
The CDF of tc is shown in Fig. 3 for different MUE where δ1 = βP wN0
tψ
ru,k (x) sin θ̂ . Moreover, δ2 =
distances from the serving SBS. Fig. 3 shows that as the wru,k (x) sin θ̂Pck (Nk , θk )/vu tc , and θ̂ = θu − θ0 + θk . For
MUE is closer to the SBS, tc takes smaller values with higher (a) to hold, we set α = 2 which is a typical value for the path
probability which is expected, since the MUE will traverse a loss exponent of LoS mmWave links [27].
shorter distance to cross the mmWave beam.
Proof. See Appendix C. Thus far, the provided analysis are focused on studying the
B. Achievable gains of caching for mobility management caching opportunities for the mobility management in single-
MUE scenarios. However, in practice, the SBSs can only serve
From (3), (4), and (17), we can find dc (u, k) which is the a limited number of MUEs simultaneously. Therefore, an HO
distance that MUE u can traverse, while using the cached decision for an MUE is affected by the decision of the other
video content. On the other hand, by having the average inter- MUEs. In this regard, we propose a cache-enabled mobility
cell distances in a HetNet, we can approximate the number of management framework to capture the inter-dependency of HO
SBSs that an MUE can pass over distance dc (u, k). Hence, the decisions in dynamic multi-MUE scenarios.
average number of SBSs that MUE is able to traverse without
performing cell search for HO is V. DYNAMIC M ATCHING FOR C ACHE - ENABLED M OBILITY
M ANAGEMENT
E [dc (u, k)]

η≈ , (18) Within the proposed mobility management scenarios, the
l
MUEs have a flexibility to perform either a vertical or horizon-
where the expected value is used, since dc (u, k) is a random tal HO, while moving to their chosen target cell. Additionally,
variable that depends on θu . Moreover, l denotes the average as elaborated in Section IV, caching enables MUEs to skip a
inter-cell distance. Here, we note that certain HO, depending on the cache size Ω. In fact, there are
Z ∞ three HO actions possible for an arbitrary MUE that is being
E [dc (u, k)] =

1 − Ftcu (vu t) dt, (19) served by an SBS: 1) Execute an HO for a new assignment
0
with a target SBS, 2) Use the cached content and mute HO,
where Ftc (.) is derived in Lemma 1. We note that (19) is the 3) Perform an HO to the MBS. Similarly, an MUE assigned
direct result of writing an expected value in terms of CDF. to the MBS can decide whether to handover to an SBS, use
Based on the definition of η in (18) and considering that the cached content, or stay connected to the MBS.
inter-frequency energy consumption linearly scales with the Our next goal is to maximize possible handovers to the
number of scans, we can make the following observation. SBSs in order to increase the traffic offload from the MBS,
Remark 1. The proposed caching scheme will reduce the subject to constraints on the HOF, SBSs’ quota, and limited
average energy consumption E s for inter-frequency cell search cache sizes. With this in mind, our goal is to find an HO policy
by a factor of 1/η with η being defined in (18). ζ for MUEs and target BSs4 that satisfies:
X
Furthermore, from the definition of γHOF in (5), we can argmin ζ(u, k0 ), (22a)
ζ u∈U
define the probability of HOF as P(Du,k < vu tMTS ) [38], !
where Du,k = tu,k /vu , and tu,k is the ToS. To compute
P
X
the HOF probability, we use the probability density function s.t. ζ(u, k)Du,k < vu tMTS ≤ Puth , (22b)
k∈K
(PDF) of a random chord length within a circle with radius a, " #
as follows: X Ωu
1− ζ(u, k) Ts ≤ , (22c)
2 Q
fD (D) = √ , (20) k∈K0
π 4a2 − D2
X
ζ(u, k) ≤ 1, (22d)
where (20) relies on the assumption that one side of the chord k∈K0
is fixed and the other side is determined by choosing a random
X
ζ(u, k) ≤ Ukth , ∀k ∈ K, (22e)
θ ∈ [0, π]. This assumption is in line with our analysis as u∈U
shown in Fig. 2. Using (20), we can find the probability of ζ(u, k) ∈ {0, 1}, (22f)
HOF as follows:
Z vu tMTS where K0 = K ∪ {k0 } and ζ is a vector of binary elements
P(Du,k < vu tMTS ) = p
2
dD ζ(u, k) ∈ {0, 1}. In fact, a variable ζ(u, k) = 1, if MUE u
0 π 4a2k − D2 is chosen to execute an HO to the target cell k, otherwise,

2 vu tMTS ζ(u, k) = 0. Constraints (22b)-(22f) must hold for all u ∈ U.
= arcsin . (21)
π 2ak In fact, the objective in (22a) is to minimize the number
In fact, γHOF is a binomial random variable whose probability of MUEs associated with MBS k0 . (22b) Pensures that once
of success depends on the MUE’s speed, cell radius, and tMTS . an MUE u is assigned to an SBS, i.e. k∈K ζ(u, k) = 1,
Hence, by reducing the number of HOs by a factor of 1/η, the the probability of HOF must be less than a threshold Puth ,
proposed scheme will reduce the expected value of the sum determined based on the QoS requirement of the MUE u’s
P
γHOF , taken over all SBSs that an MUE visits during the 4 For brevity, if not specified, we refer to a base station (BS) as either an
considered time T . SBS k ∈ K or the MBS k0 .
service. Constraint (22c) ensures that if an MUE u is not A. Handover as a matching game: Preliminaries
assigned to any SBS nor the MBS, there will be enough cached Matching theory is a mathematical framework that provides
video segments for the next Ts time duration. Moreover, polynomial time solutions for combinatorial assignment prob-
constraints (22d) and (22e) indicate, respectively, that each lems such as (22a)-(22f) [20]. In a static form, a matching
MUE can be assigned to at most one BS and each SBS can game is defined as a two-sided assignment problem between
serve maximum Ukth MUEs simultaneously. two disjoint sets of players in which the players of each set
We note that are interested to be matched to the players of the other set,
using (21), we can rewrite (22b) as
2 vu tMTS according to their preference profiles. A preference profile for
ζ(u, k) ≤ Puth , which is a linear
P
k∈K π arcsin 2ak player i, denoted by i , is defined as a complete, reflexive,
constraint. Hence, the posed problem in (22a)-(22f) is an and transitive binary relation between the elements of a given
integer linear programming (ILP), and thus, it is NP-hard. set. Within the context of our proposed cache-enabled HO
Although an approximation algorithm [39] can be employed problem, we define the matching problem as follow:
to solve (22a)-(22f), centralized algorithms are not scalable
and typically introduce latency which is not desired for Definition 1. Given the two disjoint sets of MUEs and BSs,
real-time applications such as streaming for mobile users. respectively, in U and K0 = K ∪ {k0 }, a single-period HO
Moreover, these solutions will typically rely on the current matching is defined as a many-to-one mapping µ : U ∪ K0 →
network instances, such as the location, speed and cache size U ∪ K0 that satisfies:
of the MUEs, and, hence, they fail to capture the dynamics of 1) ∀u ∈ U, µ(u) ∈ K0 ∪ {u}. In fact, µ(u) = k means u is
the system. To show this, we consider two critical scenarios, assigned to k, and µ(u) = u indicates that the MUE u
shown in Fig. 4, as follows: is not matched to any BS, and thus, will use the cached
content.
Illustrative Scenario 1: Consider a feasible solution for
2) ∀k ∈ K0 , µ(k) ⊆ U ∪ {k}, and ∀k ∈ K, |µ(k)| ≤ Ukth . In
(22a)-(22f), where an MUE u is not assigned to the target SBS
fact, µ(k) = k implies that no MUE is assigned to the
k and will use the cached content for the next Ts time duration,
BS k.
as shown in scenario 1 of Fig. 4. However, the MUE has to
3) µ(u) = k, if and only if u ∈ µ(k).
be assigned to the MBS after Ts , since eventually no target
SBS is detected. Alternatively, the MUE could be assigned Note that, by definition, the matching game satisfies con-
to k initially and fill up the cache, while later, it could use straints (22d)-(22f). More interestingly, the matching frame-
the saved cached content to reach the next target cell without work allows defining relevant utility functions per MUE and
requiring to be assigned to the MBS. SBSs, which can capture the preferences of MUEs and SBSs.
In this regard, the utility that an arbitrary MUE u ∈ U assigns
Illustrative Scenario 2: Consider a feasible solution for to an SBS k ∈ K will be:
(22a)-(22f) which assigns an arbitrary MUE 1, in Fig. 4, to a !
Φ(u, k) = Pu − P
target SBS k1 . If there are not enough cached contents for the th
X
ζ(u, k)Du,k < vu tMTS
MUE 1 to move to the next SBS and at the same time HO
k∈K
fails, the MUE has to be assigned to the MBS as shown in
2 vu tMTS
scenario 2 of Fig. 4. Alternatively, we could assign MUE 2 = Puth − arcsin . (23)
with a large cache size to the SBS k1 such that in case of an π 2ak
HOF, the MUE 2 can reach the next target SBS k2 by using Here, we observe that the utility in (23) is larger for SBSs
its available cached contents. having a larger cell radius ak . In addition, as the speed of the
MUEs increases, the utility generated from those MUEs being
These examples show that taking into account the future
assigned to an SBS decreases. Meanwhile, the utility that an
network information, such as the estimated distance from the
SBS k assigns to an MUE u is given by
next target SBS, is imperative to effectively maximize the
traffic offloads from the MBS. Therefore, an efficient HO Ωu
Γ (u, k) = Ts − . (24)
policy must take into account post-handover scenarios that Q
may occur due to the HOFs. In this regard, we propose a
In fact, an SBS assigns higher utility to MUEs that are not
framework based on dynamic matching theory [40] which
capable of using caching for the next time duration Ts . Based
allows effective mobility management, as presented in (22a)-
on the defined utility functions, the preference profile of an
(22f), while capturing future network instances, such as the
arbitrary MUE u, u , will be:
cache size, MUEs’ trajectory, and the topology of the network.
Next, we present the fundamentals of matching theory and k u k 0 ⇔ Φ(u, k) > Φ(u, k 0 ), (25a)
explain how the proposed problem can be formulated as a u u k ⇔ Φ(u, k) < 0, (25b)
dynamic matching problem.
Fig. 4: Two dynamic HO scenarios for cache-enabled mobile users.
where k u k 0 implies that SBS k is strictly more preferred Algorithm 1 DA Algorithm for Single-period Association
than SBS k 0 by MUE u. Moreover, u u k means that Between MUEs and SBSs
an SBS k is not acceptable to an MUE u, if and only if Inputs: Π , (U ∪ K, u , k ).
the assigned utility is negative. In fact, (25b) known as an Outputs: Stable matching µ∗ .
individual rationality constraint and is in line with satisfying 1: If not already accepted by an SBS, each unmatched MUE u ∈ U
applies for its most preferred SBS k u u. Remove k from u’s
the feasibility condition in (22b). Similarly, we can define the preference profile u .
preference profile of an SBS k, k , as follows 2: Each SBS k ∈ K receives the proposals from the applicants in
Step 1, tentatively accepts Ukth of most preferred MUEs from new
u k u0 ⇔ Γ (u, k) > Γ (u0 , k), (26a) applicants and the MUEs that are so far accepted in µ(k), and
k k u ⇔ Γ (u, k) < 0, (26b) rejects the rest.
3: repeat Steps 1 to 2
where (26b) is the individual rationality requirement for SBSs 4: until Each MUE u is accepted by an SBS, or u is applied for
which is equivalent to satisfying the feasibility constraint in all SBSs k u u.
5: if ∃u ∈ U , µ(u) ∈ / K and Ωu /Q < Ts , then
(22c). With this in mind, the proposed matching game is 6: µ(u) = u,
formally defined as a tuple Π , (U ∪ K, u , k ), where 7: else
u = {u }u∈U and k = {k }k∈K . 8: Assign u to the MBS.
To solve this game, one desirable solution concept is to find 9: end if
a two-sided stable matching between the MUEs and SBSs, µ∗ ,
which is defined as follow [41]:
to achieve stability for dynamic settings, such as in Scenarios
Definition 2. An MUE-SBS pair (u, k) ∈ / µ is said to be a 1 and 2, we need to incorporate the post-HO scenarios into
blocking pair of the matching µ, if and only if k u {µ(u), u} the matching game, such that no MUE can block the stability
and u k {µ(k), k}. Matching µ is stable, µ ≡ µ∗ , if there is even after experiencing an HOF. To this end, we extend the
no blocking pair. notion of one-stage stability in Algorithm 1 into a dynamic
A two-sided stable association between MUEs and SBSs stability concept that is suitable for the problem at hand.
ensures fairness for the MUEs. That is, if an MUE u envies B. Dynamic matching for mobility management in heteroge-
the association of another MUE u0 , then u0 must be preferred neous networks
by the SBS µ∗ (u0 ) to u, i.e., the envy of MUE u is not justified. To account for possible scenarios that may occur after HO,
we consider a two-stage dynamic matching game that incor-
Remark 2. For a given single-period HO matching game
porates within the preference profiles, some of the possible
Π, the deferred acceptance (DA) algorithm [20], presented
scenarios that may face the MUEs and base stations after
in Algorithm 1, is guaranteed to find a two-sided stable
handover execution. Such a dynamic matching will allow the
association µ∗ between MUEs and SBSs.
MUEs to build preference profiles over different association
Unfortunately, the DA algorithm is not suitable to capture plans rather than SBSs. An association plan is defined as a
the dynamics of the system which arise from the mobility sequence of two matchings for a given MUE or SBS. For
of the MUEs. In fact, the preference profiles of the MUEs example, kk 0 is an association plan that indicates an MUE
and SBSs only depend on the current state of the system, will be assigned to the SBS k followed by another HO to
such as the location of the MUEs, and the cache sizes. In SBS k 0 . In this regard, k1 k2 u k10 k20 means that MUE
addition, the DA algorithm cannot guarantee stability, if the u prefers plan k1 k2 to k10 k20 . With this in mind, we can
preference of the MUEs change after HOFs. Thus, to be able modify the one-period matching in Definition 1 to a relation
µ† : U ∪ K0 → (U ∪ K0 )2 , such that µ† (u) = (µ1 (u), µ2 (u)), is one of the possible ex ante stable matchings. However,
where µ1 and µ2 are one-period matchings. For example, this matching is not dynamically stable. That is because
µ† (u) = (k, u) indicates that MUE u will first perform an conditioned to µ1 (u) = k1 , the MUE-MBS pair (u1 , k0 )
HO to SBS k, µ1 (u) = k, followed by using the content of will period-2 block the matching, since k1 k0 u1 k1 u1 and
the cache after exiting the coverage of SBS k, µ2 (u) = u. k0 u1 k0 k0 k0 . In practice, such a blocking occurs if the
Next, we use the following definitions to formally define the MUE experiences an HOF with its first matching to k1 .
stability in dynamic matchings [40]:
Next, we propose an algorithm that finds a dynamically sta-
Definition 3. An MUE-BS pair (u, k) can period-1 block the ble solution for the proposed mobility management problem.
matching, if any of the following conditions is satisfied: 1)
kk u µ† (u) and uu k µ† (k); 2) ku u µ† (u) and uk k C. Dynamically stable matching algorithm for mobility man-
µ† (k); 3) uk u µ† (u) and ku k µ† (k); or 4) uu u µ† (u) agement
and kk k µ† (k). A matching is ex ante stable, if it cannot To find the dynamically stable solution, we note that the
be period-1 blocked by any MUE/BS or MUE-BS pair. solution must first admit the ex ante stability. Therefore, we
In a dynamic matching problem, either the MUEs or the propose an algorithm, inspired from [40] that yields an ex
BSs may block the matching, after knowing the outcome of ante stable association in the first stage, followed by a simple
the first matching µ1 . In this regard, we define the notion of modification to resolve any possible period-2 blocking cases.
period-2 blocking and dynamic stability as follows: For each MUE u, let Pu = ∪k∈K {kk, uk, ku} be the set of
all plans considered by u. The algorithm proceeds as follows:
Definition 4. An MUE u can period-2 block a matching µ† if Stage-1 (Finding an ex ante stable matching):
(µ1 (u), u) u µ† (u). Similarly, an MUE-BS pair (u, k) can 1) For each MUE u ∈ U, if uu u κ, for all κ ∈ Pu , then
period-2 block if any of the following conditions is satisfied: u does not send any plan proposal to the BSs. Otherwise,
1) (µ1 (u), k) u µ† (u) and (µ1 (k), u) k µ† (k), or 2) MUE u sends a plan proposal to a BS, according to the
(µ1 (u), u) u µ† (u) and (µ1 (k), k) k µ† (k). A matching most preferred plan κ∗u as follows. If κ∗u = kk, MUE u
is said to be dynamically stable, if it cannot be period-1 or sends a request for a two-period association to the BS
period-2 blocked by any MUE or MUE-BS pair5 . k. If κ∗u = ku, the MUE sends an association request to
From Definitions 3 and 4, we can see that, any dynamically BS k, only for period-1. Similarly, if κ∗u = uk, the MUE
stable matching is also an ex ante stable matching. However, sends an association request to k only for period-2. The
ex ante stability does not guarantee dynamic stability. For MUE removes κ∗u from its preference profile for the rest
example, if µ† (u) = (k, u) for an MUE u, ex ante stability of the procedure.
does not guarantee that the MUE commits to use the cache, 2) Each SBS k ∈ K receives the plan proposals and
if the first handover to SBS k fails. In other words, the MUE tentatively accepts the most preferred plans, such that
may block an ex ante stable matching after the actual outcome the quota Ukth is not violated at each period. Clearly, any
of the first matching is known. To help better understand the accepted plan κ by SBS k satisfies κ k kk.
stability for dynamic matchings, we consider the following 3) MUEs with rejected plans apply in the next round, based
simple example. on their next most preferred plan. The first stage of the
algorithm converges, once no plan is rejected.
Example 1. Consider a dynamic matching game Π † , com-
posed of MUEs U = {u1 , u2 }, MBS k0 , and SBSs K = Proposition 1. Stage-1 of the proposed algorithm in Algo-
{k1 , k2 }, with Ukth = 1 for k = k1 , k2 , as shown in scenario 2 rithm 2 converges to an ex ante stable association between
of Fig. 4. The preference plans of MUEs, MBS k0 , and SBSs MUEs and BSs.
are as follows: Proof. See Appendix D.
u1 : k1 k0 , k1 u1 , u1 k0 , u1 u1 ; u2 : k1 u2 , u2 k2 , u2 u2 ;
To avoid period-2 blockage, we introduce a certain structure
k1 : u1 k1 , u2 k1 , k1 k1 ; k2 : k2 u2 , k2 k2 ; k0 : k0 u1 , k0 k0 ;
to the preference profile of the SBSs as follows. For any SBS
where the preference profiles are sorted in descending order for whom the maximum quota of Ukth MUEs are assigned, i.e.
and association plans that are not included do not meet the |µ†2 (k)| = Ukth ,
individual rationality constraint. Here, the underlined matching µ† k µ†1 (k), µ˜2 † (k) ∪ {u} , (27)
5 In general, a matching is dynamically stable for any time t, if it cannot where µ˜2 † (k) is µ2 † (k) with one associated MUE removed
be period-t blocked by any MUE or MUE-BS pair. Extending the dynamic to accommodate a new matching with MUE u. In fact, (27)
matching to more than two periods depends on how much information is
available for MUEs about the network. In this work, we focus on a two- implies that an MUE cannot period-2 block the matching with
period matching problem, since it is more tractable and practical. any SBS k that is associated to Ukth MUEs. In addition,
Algorithm 2 Proposed Algorithm for Dynamic Matching TABLE II: Simulation parameters
Notation Parameter Value
Between MUEs and BSs fc Carrier frequency 73 GHz [27]
Inputs: Preference plans κ for all MUEs, MBS, and SBSs. Pt,k Total transmit power of SBSs [20, 27, 30] dBm
Outputs: Dynamically stable matching µ∗ . K Total number of SBSs 50
w Available Bandwidth 5 GHz [27]
Phase 1:
(αLoS ,αNLoS ) Path loss exponent (2, 3.5) [27]
1: For each MUE u ∈ U, if uu u κ, for all κ ∈ Pu , then u does d0 Path loss reference distance 1 m [27]
not send any plan proposal to the BSs. Otherwise, MUE u sends Gmax Antenna main lobe gain 18 dB [28]
a plan proposal to a BS, according to the most preferred plan Gmin Antenna side lobe gain −2 dB [28]
κ∗u . Nk Number of mmWave beams 3
2: Each SBS k ∈ K receives the plan proposals and tentatively θm , θ k beam width 10◦ [28]
N0 Noise power spectral density −174 dBm/Hz
accepts most preferred plans (also compared to plans that are tMTS Minimum time-of-stay 1s [24]
previously accepted), such that the quota Ukth is not violated at Q Play rate 1k segments per second
each period. Clearly, any accepted plan κ by SBS k satisfies B Size of video segments 1 Mbits
κ k kk. (vmin , vmax ) Minimum and maximum MUE speeds (1, 16) m/s [24]
3: repeat Steps 1 to 2 Es Energy per inter-frequency scan 3 mJ [22]
4: until No plan is rejected. The yielded ex ante stable matching is
denoted by µ† = (µ†1 , µ†2 ).
Phase 2: Proof. See Appendix E.
†
5: if ∃u ∈ U, µ2 (u) = u, then apply DA algorithm in Algorithm
To analyze the signaling overhead of the proposed algo-
1 to the subset of MUEs with µ†2 (u) = u and the subset of BSs
with |µ†2 (k)| < Ukth , considering the constraints in (27) and (28).
rithm, we consider the total number of HO requests sent to a
Return yielded matching. target SBS by the MUEs. This metric allows to capture the
6: else overhead of the proposed scheme, in terms of the number of
7: return µ† . control signals that must be handled by the SBS. Additional
8: end if control signals from the SBSs to MUEs can be managed by
using a broadcast channel and do not significantly contribute to
! the overhead of the proposed scheme. In this regard, consider
† 2 vu tMTS
(µ1 (k0 ), u) k0 µ ⇐⇒ Pµth† (u) − arcsin < , the worst-case scenario in which the initial cache size is
1 π 2aµ† (u) Ωu = 0 for all u ∈ U. Therefore, all MUEs seek to perform an
1
(28) HO to the target SBS k by sending a request for plan κ = ku
during Stage-1 of the proposed algorithm. The SBS k accepts
where is a non-negative scalar. In fact, (28) allow MUEs
up to Ukth association plans and rejects the rest. Clearly, if there
that are assigned to SBSs in period 1, with not small enough
is one target SBS for the MUEs, the signaling overhead will
HOF probability, to be assigned to the MBS in period 2.
be O(U ). Otherwise, rejected MUEs will send an HO request
Another alternative was to set P th a small value from the start.
to the next target SBS, based on their preference profiles. The
However, this policy will discourage MUEs to be assigned
maximum signaling overhead occurs for a case when all MUEs
to SBSs and could increase the load on the MBS. With this
have the same preference profile as it introduces the highest
in mind, we construct the second stage of the algorithm as
competition among MUEs. In this case, the signaling overhead
follows:
of the proposed algorithm will be O(U K). In addition, in
Stage-2 (Remove period-2 blocking pairs): Apply the de-
Section VI, we will discuss how caching capabilities will
ferred acceptance algorithm shown in Algorithm 1 to a subset
reduce the overhead of the proposed algorithm.
of MUEs with µ†2 (u) = u, and subset of BSs with |µ†2 (k)| <
Ukth , while considering the constraints in (27) and (28). The VI. S IMULATION R ESULTS
proposed two-stage algorithm is summarized in Algorithm 2. For simulations, we consider a HetNet composed of K = 50
Reconsidering Example 1, it is easy to follow that Algorithm SBSs distributed uniformly across a circular area with radius
2 yields the following solution which is dynamically stable6 : 500 meters [24] with a single MBS located at the center and
u1 : k1 k0 , k1 u1 , u1 k0 , u1 u1 ; u2 : k1 u2 , u2 k2 , u2 u2 ;
a minimum inter-cell distance of 30 meters [42]. Moreover,
k1 : u1 k1 , u2 k1 , k1 k1 ; k2 : k2 u2 , k2 k2 ; k0 : k0 u1 , k0 k0 . the transmit power of SBSs are chosen randomly from the
For the proposed algorithm, we can state the following results: set of powers in [20, 27, 30] dBm. The main parameters are
summarized in Table I. In our simulations, we consider the
Theorem 3. The proposed two-stage algorithm in Algorithm overall transmit-receive antenna gain from an interference link
2 is guaranteed to converge to a dynamically stable association to be random. All statistical results are averaged over a large
between MUEs and BSs. number of independent runs. Next, we first investigate the
6 Here, we assume that (28) holds for u . Otherwise, the ex ante stable
gains achievable by the proposed cache-enabled scheme for a
1
solution in Example 1 is also dynamically stable, since k0 will not make a single user scenario. Then, we will evaluate the performance
period-2 block pair with u1 . of the proposed dynamic matching approach by extending the
0.25 0.8
Proposed scheme with caching Conventional HO without caching
Conventional HO without caching Proposed algorithm, Ωu = 0 segments
0.7
Proposed algorithm, Ωu = 104 segments
0.2
0.6
Average handover failure
Average HOF probability

0.5
0.15
0.4
0.1
0.3
0.2
0.05
0.1
0 0
3 10 30 45 60 2 4 6 8 10 12
MUE speed (km/h) MUE speed (m/s)
Fig. 5: HOF vs. different MUE speeds. Fig. 7: Average HOF probability vs. MUEs’ speeds.
2
10
Fig. 6 shows the achievable rate of caching for an MUE
with vu = 60 km/h, as a function of different initial distances
10
1
ru,k (x) for various θu . The results in Fig. 6 show that even
for MUEs with high speeds, the achievable rate of caching
10
0 is significant, exceeding 10 Gbps, for all θu values and inital
Achievable rate (Gbps)
distance of 20 meters from the SBS. However, we can observe

that the blockage can noticeably degrade the performance. In
−1 NLoS, θu = 0.1π
10
NLoS, θu = 0.2π
fact, for NLoS scenarios, the maximum achievable rate at a
NLoS, θu = 0.3π distance of 20 meters decreases to 2 Gbps.
NLoS, θu = 0.4π
−2
10 NLoS, θu = 0.5π B. Performance of the proposed dynamically stable mobility
LoS, θu = 0.1π
LoS, θu = 0.2π management algorithm
−3 LoS, θu = 0.3π
10
LoS, θu = 0.4π
Here, we consider the set of MUEs entering a target cell
LoS, θu = 0.5π coverage region with random directions and speeds. Moreover,
10
−4 the cache sizes of the MUEs are initially Ωu = 104 segments
20 40 60 80 100 120 140
MUE distance from SBS (m)
160 180 200
for all MUEs. In addition, each SBS can serve up to Ukth = 10
Fig. 6: Achievable rate of caching vs. ru,k (x) for different θu . MUEs. Depending on the speed of the MUE, its direction,
and the location of the next target SBS, MUEs form their
results for scenarios with multiple MUEs in which SBSs can preferences over different plans as elaborated in Section V.
only serve a limited number of MUEs. In Fig. 7, the average HOF probability of the proposed
algorithm is compared with a conventional scheme that does
A. Analysis of the proposed cache-enabled mobility manage- not incorporate caching, versus the speed of the MUEs. The
ment for single user scenarios HOF probability is defined as the ratio of the MUEs with
Fig. 5 compares the average HOF of the proposed scheme HOF to the total number of MUEs, for U = 20 and Ukth = 10.
with a conventional HO mechanism without caching. This The results in Fig. 7 show that the HOF probability increases
baseline scheme is based on the same system as the proposed with the speed of the MUEs, since the ToS will decrease for
approach, but it does not incorporate the caching mechanism higher MUE speeds. In addition, we observe that the proposed
developed in Section V. The results clearly demonstrate that algorithm can significantly reduce the HOF probability by
caching capabilities, as proposed here, will significantly im- leveraging the information on the MUE’s trajectory and the
prove the HO process for dense HetNets. In fact, the results network’s topology. Fig. 7 also shows that for a non-zero
in Fig. 5 show that caching over mmWave frequencies will initial cache sizes of Ωu = 104 segments, the algorithm is
reduce HOF for all speeds, reaching up to 45% for MUEs considerably robust against HOF. In fact, the HOF probability
with vu = 60 km/h. declines for speeds beyond vu = 8 m/s, since higher speed
10 120
vu = 8 m/s vu = 8 m/s
Inter−frequency measurement energy saving (mJ)

9
vu = 10 m/s vu = 10 m/s
vu = 12 m/s 100
8 vu = 12 m/s
7
Load of the target SBS
80
6
5 60
4
40
3
2
20
1
0 0
10 15 20 25 30 35 40 45 50 10 20 30 40 50
Number of MUEs Number of MUEs
Fig. 8: Load of the target SBS vs. the number of MUEs. Fig. 9: Energy savings for inter-frequency measurements vs number
of MUEs.
allows the MUE to traverse larger distance before the cached

45
video segments run out. Therefore, more MUEs will be able
to skip an HO to the target cell and use the cached content to 40
vu = 2 m/s
vu = 6 m/s
move to the next available SBS.
vu = 8 m/s
35
Fig. 8 shows the load of the target cell versus the number of
MUEs for different MUE speeds vu = 8, 10, and 12 m/s, SBS 30
Signaling overhead
quota Ukth = 10, and initial cache size Ωu = 104 segments.

Here, we observe that the proposed algorithm associates less 25
MUEs to the target cell as the speed increases. That is due to

20
two reasons: 1) higher speeds decrease the ToS and increase
the chances of HOFs, and 2) with higher speeds, MUEs can 15
traverse longer distances by using Ωu cached segments and
10
it is more likely that they can reach to the next target SBS.
Fig. 8 shows that the load of the target cell reduces up to 45% 5
when vu increases from 8 to 10 m/s for U = 40 MUEs.
0
In Fig. 9, the inter-frequency measurement energy savings 10 15 20 25 30 35 40 45 50
yielded by the proposed algorithm are shown as a function of Number of MUEs
Fig. 10: Signaling overhead vs. number of MUEs.
the number of MUEs. Fig. 9 shows the total saved energy for
MUEs that will use the cached content and do not perform any
inter-frequency measurements for handover to an SBS for an are consistent with those shown in Fig. 8. In fact, as the speed
initial cache size of Ωu = 104 segments and different MUE of MUEs increases, the HOF probability increases, and thus,
speeds. For U = 50, MUEs that perform conventional han- MUEs tend to be assigned to the MBS or use their cached
dover without caching will require U E s = 150 mJ total energy content. In addition, fast moving MUEs are more likely to
for performing inter-frequency measurements. However, the reach the next target cell before the cached content runs out.
results in Fig. 9 show that the proposed scheme achieves up In Fig. 10, we show the signaling overhead resulting from
to 80%, 52%, and 29% gains in saving energy, respectively, the proposed algorithm versus the number of MUEs, for
for MUE speeds vu = 8, 10, and 12 m/s by leveraging cached Ωu = 104 initial cache size and different MUE speeds.
segments and muting unnecessary cell search. Given that the Here, we refer to the signaling overhead as the number of
required energy for measurements linearly scales with the HO requests sent to the target SBS by the MUEs. Fig. 10
number of MUEs, the results in Fig. 9 can also be interpreted shows that for low speeds vu = 2 m/s, almost all MUEs
as the offloading gains of the proposed approach, compared will attempt to hand over to the target SBS, since the time
with conventional HO with no caching. Moreover, these results needed for traversing the SBS coverage is longer than the time
available by using the cached content. Nonetheless, the results B. Proof of Lemma 1
in Fig. 10 clearly demonstrate that the proposed algorithm has From (9), Ftc (t0 ) = P(ruc ≤ vu t0 ). To find this prob-
a manageable overhead, not exceeding 17 requesting signals ability, we note that ruc ≤ vu t0 if MUE moves between
for a network size of U = 50 with vu = 8 m/s. In fact, it is two line segments of length vu t0 that connect MUE to line
interesting to note that although mobility management is, in y = x cos θ0 . Depending on ru,k (x), the intersection of line
general, more challenging for high speed MUEs, the overhead segment with y = x cos θ0 may have one or two solutions.
of the proposed algorithm decreases for high speed scenarios. In case of two intersection points, the two line segments
This is due to the fact that high speed MUEs use the cached will make two equal angles with the perpendicular line from
content more effectively than slow-moving MUEs. xu , to y = x cos θ0 , which each is obviously
min equal to
VII. C ONCLUSIONS r
π−(π/2−θk )− θ̂ = π/2+θk − θ̂ = arccos vuut0 . Therefore,
In this paper, we have proposed a comprehensive framework
min
for mobility management in integrated microwave-millimeter 2 ru
wave cellular networks. In particular, we have shown that Ftc (t0 ) = arccos . (31)
π − θk vu t0
by smartly caching video contents while exploiting the dual-
In fact, θu must be within a range of π − θk for ruc ≤ vu t0
mode nature of the network’s base stations, one can provide
to be valid. Now, if this angle is greater than π/2 − θk , only
seamless mobility to the users. We have derived various
one intersection point exists. Equivalently,
fundamental results on the probability and the achievable rate
min min
for caching video contents by leveraging millimeter wave 1 ru ru
high capacity transmissions. In addition, to capture the dy- Ftc (t0 ) = arccos +arccos . (32)
π − θk vu t0 ru,k (x)
namics of the mobility management, we have formulated the
Integrating (31) and (32), the CDF for caching duration can
multi-user handover problem as a dynamic matching game
be written as (13).
between the mobile users and small base stations. To solve this
game, we have proposed a novel algorithm that is guaranteed C. Proof of Theorem 2
to converge to a dynamically stable handover mechanism.
Theorem 1 implies that with probability 1 − Pck (Nk , θk ),
Moreover, we have shown that the proposed cache-enabled
only µW coverage is available for an MUE. Therefore, the
mobility management framework provides significant gains in
average achievable rate for caching over the mmWave fre-
reducing the number of handovers, energy consumption for
quencies is given by (15). To simplify (15), we have
inter-frequency scanning, as well as mitigating the handover
failure. Numerical results have corroborated our analytical ru,k cos θ = ru,k (x) + ru cos θ̂, ru,k sin θ = ru sin θ̂, (33)
results and showed that the significant rates for caching can be
achieved over the mmWave frequencies, even for fast mobile where θ̂ = θu − θ0 + θk and θ is an angle between the line
users. In addition, the results have shown that the proposed connecting MUE to SBS, ranging from 0 to θk . Moreover, ru
approach substantially decreases the handover failures and is the current traversed distance, with ru = ruc once the MUE
provides significant energy savings in heterogeneous networks. reaches x0 by the end of caching duration, as shown in Fig.
A PPENDIX 2. From (33), we find ru,k = ru,k (x) sin θ̂/ sin(θ̂ − θ). By
A. Proof of Theorem 1 changing the integral variable ru to θ, we can write (15) as
Due to the equidistant beams, we have Z θk cos(θ̂ − θ)
1˜ 1h i 1
2π
R̄c (u, k) = δ2 log 1 + δ1 sinα (θ̂ − θ) dθ,
AoB = AB =
[ 2π − AoB =
¯ 2π − − θk 0 sin2 (θ̂ − θ)
2 2 2 Nk (34)

1 θk
= 1− π+ . (29) where δ1 = βPt ψ(ru,k (x) sin θ̂)−α /wN0 and δ2 =
Nk 2
wru,k (x) sin θ̂Pck (Nk , θk )/vu tc . Next, we can directly con-
Given that an arbitrary MUE can enter the circle in Fig. 2 at clude (16) from (34) by substituting f (θ) = sin(θ̂ − θ) in
any direction, this MUE will be instantly covered by mmWave
with probability P(x ∈ A) = N2π , where A ⊂ R2 denotes
k θk
(34). For α = 2, which is a typical value for the path loss
exponent for LoS mmWave links, (16) can be simplified into
the part of circle’s perimeter that overlaps with mmWave (17) by taking the integration by parts in (16).
beams. Therefore,
1 [ D. Proof of Proposition 1
Pck (Nk , θk ) = P(x ∈ A) + [1 − P(x ∈ A)] AoB, (30)
2π Assume an MUE-BS pair (u, k) period-1 blocks the match-
where (30) results from the fact that θ ∼ U [0, 2π]. Therefore, ing µ† . In consequence, there is a plan κ ∈ {ku, uk, kk} for u
from (29) and (30), the probability of crossing a mmWave and a corresponding plan for k that both prefer to their current
beam follows (7). matching in µ† . If κ u µ† (u), then the MUE u must have
sent a proposal for κ to k prior to its associated plan in µ† . [9] H. Zhang, N. Meng, Y. Liu, and X. Zhang, “Performance evaluation for
Since κ is not eventually accepted, that means at some point, local anchor-based dual connectivity in 5G user-centric network,” IEEE
Access, vol. 4, pp. 5721–5729, September 2016.
the SBS k has rejected κ in favor of another plan. Since the [10] I. Elgendi, K. S. Munasinghe, and A. Jamalipour, “Mobility management
matching for SBSs improves at each round, we conclude that in three-tier sdn architecture for densenets,” in 2016 IEEE Wireless
Communications and Networking Conference, pp. 1–6, April 2016.
κ is less preferred by k compared to µ† (k). This contradicts [11] S. G. Park and Y. S. Choi, “Mobility enhancement in centralized
the first assumption, thus, such a period-1 blocking pair does mmwave-based multi-spot beam cellular system,” in 2015 International
not exist and µ† is ex ante stable. Conference on Information and Communication Technology Conver-
gence, pp. 200–205, October 2015.
[12] J. Qiao, X. S. Shen, J. W. Mark, and L. Lei, “Video quality provisioning
E. Proof of Theorem 3 for millimeter wave 5G cellular networks with link outage,” IEEE
Transactions on Wireless Communications, vol. 14, pp. 5692–5703,
From Proposition 1, the solution is guaranteed to be ex ante October 2015.
stable. Therefore, MUEs and BSs will not period-1 block the [13] O. Semiari, W. Saad, M. Bennis, and B. Maham, “Mobility management
for heterogeneous networks: Caching meets millimeter wave to provide
matching. The rest of the proof easily follows the fact that the seamless handover,” in Proc. of the 2017 IEEE GLOBECOM, December
BSs will not make a period-2 blocking pair with any MUE, 2017.
due to the constraints in (27) and (28). In fact, if there is any [14] X. Wang, M. Chen, T. Taleb, A. Ksentini, and V. C. M. Leung,
“Cache in the air: exploiting content caching and delivery techniques
period-2 blocking pair (u, k), there are four possible cases to for 5G systems,” IEEE Communications Magazine, vol. 52, pp. 131–
consider: 1) kk u µ∗ (u) and uu k µ∗ (k), 2) uk u µ∗ (u) 139, February 2014.
[15] Y. Rao, H. Zhou, D. Gao, H. Luo, and Y. Liu, “Proactive caching
and ku k µ∗ (k), 3) uk u µ∗ (u) and u0 u k µ∗ (k), where for enhancing user-side mobility support in named data networking,”
u0 6= u, or 4) k 0 k u µ∗ (u) and u0 u k µ∗ (k), where k 0 6= k in Proc. of 2013 International Conference on Innovative Mobile and
and u0 6= u. The first two cases are not possible, since they Internet Services in Ubiquitous Computing, pp. 37–42, July 2013.
[16] K. Poularakis and L. Tassiulas, “Exploiting user mobility for wireless
indicate that (u, k) can period-1 block µ∗ which contradicts ex content delivery,” in 2013 IEEE International Symposium on Information
ante stability. Considering the last two cases, since MUE u is Theory, pp. 1017–1021, July 2013.
[17] A. S. Gomes, B. Sousa, D. Palma, V. Fonseca, Z. Zhao, E. Monteiro,
not associated with SBS k in period-2, that implies that k has T. Braun, P. Simoes, and L. Cordeiro, “Edge caching with mobility
already been assigned to Ukth MUEs. Otherwise, u would be prediction in virtualized LTE mobile networks,” Future Generation
assigned to k during the second stage of Algorithm 2. Hence, Computer Systems, 2016.
[18] M. Chen, M. Mozaffari, W. Saad, C. Yin, M. Debbah, and C. S. Hong,
due to the constraint (27), k will not make a period-2 blocking “Caching in the sky: Proactive deployment of cache-enabled unmanned
pair with u. Similarly, the MBS will not make a period-2 aerial vehicles for optimized quality-of-experience,” IEEE Journal on
blocking pair with any MUE u that is not assigned to the MBS Selected Areas in Communications, vol. 35, pp. 1046–1061, May 2017.
[19] M. Chen, W. Saad, C. Yin, and M. Debbah, “Echo state networks
during the second stage. Thus, no period-2 blocking pair exists for proactive caching in cloud-based radio access networks with mo-
and the solution µ∗ satisfies dynamic stability. bile users,” IEEE Transactions on Wireless Communications, vol. 16,
pp. 3520–3535, June 2017.
R EFERENCES [20] D. Gale and L. Shapley, “College admissions and the stability of
marriage,” Amer. Math. Monthly, vol. 69, 1962.
[1] D. Lopez-Perez, I. Guvenc, and X. Chu, “Mobility management chal- [21] E. Jorswieck, “Stable matchings for resource allocation in wireless
lenges in 3GPP heterogeneous networks,” IEEE Communications Mag- networks,” in Proc. of 17th International Conference on Digital Signal
azine, vol. 50, pp. 70–78, December 2012. Processing (DSP), (Corfu, Greece), July 2011.
[2] I. F. Akyildiz, J. Xie, and S. Mohanty, “A survey of mobility manage- [22] A. Prasad, O. Tirkkonen, P. Lunden, O. N. Yilmaz, L. Dalsgaard,
ment in next-generation all-IP-based wireless systems,” IEEE Wireless and C. Wijting, “Energy-efficient inter-frequency small cell discovery
Communications, vol. 11, pp. 16–28, August 2004. techniques for LTE-Advanced heterogeneous network deployments,” in
[3] F. Giust, L. Cominardi, and C. J. Bernardos, “Distributed mobility IEEE Communications Magazine, pp. 72–81, May 2013.
management for future 5G networks: overview and analysis of existing [23] K. I. Pedersen, P. H. Michaelsen, C. Rosa, and S. Barbera, “Mobility
approaches,” IEEE Communications Magazine, vol. 53, pp. 142–149, enhancements for lte-advanced multilayer networks with inter-site car-
January 2015. rier aggregation,” IEEE Communications Magazine, vol. 51, pp. 64–71,
[4] A. Prasad, O. Tirkkonen, P. LundÃ n,
c O. N. C. Yilmaz, L. Dalsgaard, May 2013.
and C. Wijting, “Energy-efficient inter-frequency small cell discovery [24] 3GPP, “E-UTRA: Mobility enhancements in heterogeneous networks,”
techniques for LTE-advanced heterogeneous network deployments,” 3rd Generation Partnership Project, vol. Rel 11, September 2012.
IEEE Communications Magazine, vol. 51, pp. 72–81, May 2013. [25] A. Ravanshid, P. Rost, D. S. Michalopoulos, V. V. Phan, H. Bakker,
[5] D. Xenakis, N. Passas, L. Merakos, and C. Verikoukis, “Mobility D. Aziz, S. Tayade, H. D. Schotten, S. Wong, and O. Holland, “Multi-
management for femtocells in LTE-advanced: Key aspects and survey of connectivity functional architectures in 5g,” in 2016 IEEE International
handover decision algorithms,” IEEE Communications Surveys Tutorials, Conference on Communications Workshops, pp. 187–192, May 2016.
vol. 16, pp. 64–91, First 2014. [26] E. M. Royer, P. M. Melliar-Smith, and L. E. Moser, “An analysis of the
[6] A. Ahmed, L. M. Boulahia, and D. Gaiti, “Enabling vertical handover optimum node density for ad hoc mobile networks,” in ICC 2001. IEEE
decisions in heterogeneous wireless networks: A state-of-the-art and International Conference on Communications. Conference Record (Cat.
a classification,” IEEE Communications Surveys Tutorials, vol. 16, No.01CH37240), vol. 3, pp. 857–861 vol.3, June 2001.
pp. 776–811, Second 2014. [27] A. Ghosh, R. Ratasuk, P. Moorut, T. S. Rappaport, and S. Sun,
[7] K. Vasudeva, M. Simsek, D. Lopez-Perez, and I. Guvenc, “Impact of “Millimeter-Wave enhanced local area systems: A high-data-rate ap-
channel fading on mobility management in heterogeneous networks,” proach for future wireless networks,” IEEE Journal on Selected Areas
in 2015 IEEE International Conference on Communication Workshop, in Communications, vol. 32, no. 6, pp. 1152 –1163, June 2014.
June 2015. [28] S. Singh, M. N. Kulkarni, A. Ghosh, and J. G. Andrews, “Tractable
[8] M. Khan and K. Han, “An optimized network selection and handover model for rate in self-backhauled millimeter wave cellular networks,”
triggering scheme for heterogeneous self-organized wireless networks,” IEEE Journal on Selected Areas in Communications, vol. 33, pp. 2196–
Mathematical Problems in Engineering, vol. 16, pp. 1–11, 2014. 2211, October 2015.
[29] J. G. Andrews, T. Bai, M. N. Kulkarni, A. Alkhateeb, A. K. Gupta, Walid Saad (S’07, M’10, SM’15) received his
and R. W. Heath, “Modeling and analyzing millimeter wave cellular Ph.D degree from the University of Oslo in 2010.
systems,” IEEE Transactions on Communications, vol. 65, pp. 403–430, Currently, he is an Associate Professor at the De-
Jan 2017. partment of Electrical and Computer Engineering
[30] H. N. Dai, K. W. Ng, R. W. Wong, and M. Y. Wu, “On the capacity at Virginia Tech, where he leads the Network
of multi-channel wireless networks using directional antennas,” in IEEE Science, Wireless, and Security (NetSciWiS) lab-
INFOCOM 2008 - The 27th Conference on Computer Communications, oratory, within the Wireless@VT research group.
April 2008. His research interests include wireless networks,
[31] S. Akoum, O. E. Ayach, and R. W. Heath, “Coverage and capacity in machine learning, game theory, cybersecurity, un-
mmwave cellular systems,” in 2012 Conference Record of the Forty Sixth manned aerial vehicles, and cyber-physical systems.
Asilomar Conference on Signals, Systems and Computers (ASILOMAR), Dr. Saad is the recipient of the NSF CAREER award
pp. 688–692, Nov 2012. in 2013, the AFOSR summer faculty fellowship in 2014, and the Young
[32] A. Guo and M. Haenggi, “Asymptotic deployment gain: A simple ap- Investigator Award from the Office of Naval Research (ONR) in 2015. He
proach to characterize the sinr distribution in general cellular networks,” was the author/co-author of six conference best paper awards at WiOpt in
IEEE Transactions on Communications, vol. 63, pp. 962–976, March 2009, ICIMP in 2010, IEEE WCNC in 2012, IEEE PIMRC in 2015, IEEE
2015. SmartGridComm in 2015, and EuCNC in 2017. He is the recipient of the
[33] T. Bai and R. W. Heath, “Coverage and rate analysis for millimeter-wave 2015 Fred W. Ellersick Prize from the IEEE Communications Society. From
cellular networks,” IEEE Transactions on Wireless Communications, 2015-2017, Dr. Saad was named the Stephen O. Lane Junior Faculty Fellow
vol. 14, pp. 1100–1114, Feb 2015. at Virginia Tech and, in 2017, he was named College of Engineering Faculty
[34] M. K. Samimi and T. S. Rappaport, “Local multipath model parameters Fellow. He currently serves as an editor for the IEEE Transactions on Wireless
for generating 5g millimeter-wave 3gpp-like channel impulse response,” Communications, IEEE Transactions on Communications, IEEE Transactions
in 2016 10th European Conference on Antennas and Propagation on Mobile Computing, and IEEE Transactions on Information Forensics and
(EuCAP), pp. 1–5, April 2016. Security.
[35] V. Joseph and G. de Veciana, “Nova: Qoe-driven optimization of dash-
based video delivery in networks,” in IEEE INFOCOM 2014 - IEEE
Conference on Computer Communications, pp. 82–90, April 2014.
[36] F. Fu and M. V. D. Schaar, “A systematic framework for dynamically
optimizing multi-user wireless video transmission,” IEEE Journal on
Selected Areas in Communications, vol. 28, pp. 308–320, April 2010.
[37] V. Joseph and G. de Veciana, “Jointly optimizing multi-user rate adapta- Mehdi Bennis (Senior Member, IEEE) is an asso-
tion for video transport over wireless systems: Mean-fairness-variability ciate professor at the University of Oulu, Finland
tradeoffs,” in 2012 Proceedings IEEE INFOCOM, pp. 567–575, March and an Academy of Finland Research Fellow. He
2012. received his M.Sc. degree from the Ecole Polytech-
[38] C. H. M. de Lima, M. Bennis, and M. Latva-aho, “Modeling and analysis nique Federale de Lausanne (EPFL), Switzerland
of handover failure probability in small cell networks,” in Proc. of and the Eurecom Institute, France in 2002. He
IEEE Conference on Computer Communications Workshops, pp. 736– obtained his Ph.D. degree in electrical engineering
741, April 2014. December 2009 on spectrum sharing for future mo-
[39] P. Raghavan and C. D. Thompson, “Randomized rounding: A technique bile cellular systems. His main research interests
for provably good algorithms and algorithmic proofs,” Combinatorica, are in radio resource management, heterogeneous
vol. 7, no. 4, pp. 365–374, 1987. networks, game theory and machine learning. He
[40] S. Kadam and M. H. Kotowski, “Multi-period matching,” has published more than 100 research papers in international conferences,
Available online at http:// scholar.harvard.edu/ kadam/ publications/ journals, book chapters and patents. Dr Bennis gave a dozen of tutorials
multi-period-matching, April 2016. at IEEE flagship conferences. He was also the recipient of several prizes
[41] A. E. Roth and M. A. O. Sotomayor, Two-sided matching: A study including the 2015 Fred W. Ellersick Prize from the IEEE Communications
in game-theoretic modeling and analysis. Cambridge University Press, Society (COMSOC), the 2016 IEEE COMSOC Best Tutorial Prize, the 2017
1992. EURASIP Best Paper Award for the Journal on Wireless Communications and
[42] 3GPP, “E-UTRA: Further advancements for E-UTRA physical layer Networking (JWCN), and recently the best paper Award at EUCNC 2017.
aspects,” 3rd Generation Partnership Project, vol. Rel 9, 2010.
Behrouz Maham (S’07, M’10, SM’15) received

Omid Semiari (S’14, M’17) is an Assistant Pro- the BSc and MSc.degrees in electrical engineering
fessor at the Electrical Engineering Department at from the University of Tehran, in 2005 and 2007,
Georgia Southern University. He received the BSc respectively, and the PhD degree from the University
and MSc degrees in electrical engineering from the of Oslo, in 2010. From September 2008 to August
University of Tehran, in 2010 and 2012, respectively, 2009, he was with the Department of Electrical
and the PhD degree from Virginia Tech, in 2017. Engineering, Stanford University. He is currently an
His research interests include wireless networks, Assistant Professor with the School of Engineering,
millimeter wave communications, context-aware re- Nazarbayev University. He was a faculty with the
source allocation, matching theory, and machine School of ECE, University of Tehran, from Sep.
learning. In 2014, Dr. Semiari has worked as an 2011 to Sep. 2015. Dr. Maham is TWAS-affiliated
intern at Bell Labs, in Stuttgart, on anticipatory, and has more than 120 publications in major technical journals and confer-
context-aware resource management in cellular networks. In 2016, he has ences. His fields of interest include wireless communication and networking
joined Qualcomm CDMA Technologies (QCT) for a summer internship, and nano-neural communication systems.
working on LTE-Advanced modem design. Dr. Semiari is the recipient of
several research fellowship awards, including DAAD (German Academic
Exchange Service) scholarship and NSF student travel grant. He has actively
served as a reviewer for flagship IEEE Transactions and conferences and
participated as the technical program committee (TPC) member for a variety
of workshops at IEEE conferences, such as ICC and GLOBECOM. Currently,
he serves as a member of editorial board for the IEEE ComSoc TCBNC blog
and IEEE ComSoc Technical Committee-TCCN Newsletter.

Caching Meets Millimeter Wave Communications For

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Caching Meets Millimeter Wave Communications For

Загружено:

Авторское право:

Доступные форматы

This article has been accepted for publication in a future issue of this journal, but has not been

Caching Meets Millimeter Wave Communications for

at the µW network. ongoing IP service can be cached in advance and continuously

traffic offload from the µW frequencies.

TABLE I: Variables and notations

III. A NALYSIS OF M OBILITY M ANAGEMENT WITH

y = x tan(θ0 − θk ), y = x tan(θ0 ), x > 0. (8) 0.8

Assuming that the MUE u is currently located on the angle 0.7

side x = y cos(θ0 −θk ), as shown by the

Fig. 4: Two dynamic HO scenarios for cache-enabled mobile users.

Average HOF probability

distance of 20 meters from the SBS. However, we can observe

Inter−frequency measurement energy saving (mJ)

allows the MUE to traverse larger distance before the cached

quota Ukth = 10, and initial cache size Ωu = 104 segments.

MUEs to the target cell as the speed increases. That is due to

Behrouz Maham (S’07, M’10, SM’15) received

Вам также может понравиться