Mobile Networks and Applications

https://doi.org/10.1007/s11036-019-01421-5

A Dynamic Resource Scheduling Scheme in Edge Computing

Satellite Networks
Feng Wang 1 & Dingde Jiang 1 & Sheng Qi 1 & Chen Qiao 1 & Lei Shi 2

# Springer Science+Business Media, LLC, part of Springer Nature 2020

Abstract
The LEO satellite network has been a valuable architecture due to its characteristics of wide coverage and low transmission delay.
Utilizing LEO satellites as edge computing nodes to provide reliable computing services for access terminals will be the
indispensable paradigm of integrated space-air-ground network. However, the design of resource division strategy in edge
computing satellite (ECS) is not easy, considering different accessing planes and resource requirements of terminals.
Moreover, network topology, available resources and relative motion need to be analyzed comprehensively to establish ECS
collaborative network for emergency situations. To address these problems, a three-layer network architecture combined with
software defined networking (SDN) model is proposed to guide the inter-satellite link (ISL) connection and ECS resource
scheduling. The advanced K-means algorithm (AKG) and breadth-first-search-based spanning tree algorithm (BFST) are pro-
vided to realize ECS resource division and ISL construction respectively. Simulation results show that the proposed dynamic
resource scheduling scheme is feasible and effective.

Keywords Edge computing . LEO satellite network . Resource scheduling . Inter-satellite link . Software defined networking

1 Introduction its coverage [4]. If the communication target is not in the

coverage region, cross-domain communication can be real-
Low earth orbit (LEO) constellation network having an im- ized through the inter-satellite link (ISL). However, with the
portant role in the future space-air-ground integrated commu- development of intelligent terminals, communication data is
nication network. Compared with about 500 ms communica- no longer just voice and text, but images and video, probably
tion delay of geosynchronous earth orbit (GEO) satellites, using VR and AR technologies. Moreover, the increasing
LEO constellation has the advantages of low delay, high ca- number of special function terminals, such as unmanned aerial
pacity, full coverage and manageability [1, 2]. Moreover, it vehicles (UAV) and intelligent sensors, makes the demand for
can guarantee the efficient communication in the areas such real-time data processing impending. Data transmission be-
as polar region, desert, oceans and air that are difficult to reach tween the terrestrial command center and LEO satellite has
by GEO satellite or terrestrial base station, so as to achieve been difficult to meet the low latency computing of terminals
global network coverage [3]. Recently SpaceX, OneWeb, [5]. The main requirement is to improve the fast data comput-
LeoSat and other space Internet companies have released their ing ability of LEO satellites when massive terminals
own structure of LEO constellation, aiming to preempt near- accessing. Therefore, the fusion of LEO constellation and
earth space resources. In the LEO constellation, each satellite the edge computing paradigm has gradually attracted attention
can provide communication services for the terminals within to enhance the real-time management for intelligent terminals.
Edge computing is an emerging computing model, which
possesses sufficient computing and storage resources on the
* Dingde Jiang edge of the Internet that very close to intelligent terminals [6].
merry_99@sina.com So the data processing can be performed at the edge server,
and the terrestrial command center is more responsible for
1
School of Astronautics and Aeronautic, University of Electronic storing processing results and further big data analysis [7].
Science and Technology of China, Chengdu 611731, China The edge server can be deployed in the LEO satellite, making
2
Department of Computing, Carlow Institute of Technology, LEO satellites become edge computing satellites (ECSs),
Carlow, Ireland where the data processing module for terminals are installed.
 Mobile Netw Appl

Terminals utilize the computing resources of ECSs to realize and scheduling of resources in ECS network is a novel para-
data analysis and accept the unified scheduling of ECSs. digm. But the problem of how to make better resource alloca-
Computing resources of ECS can be split into KVM-based tion configuration and how to carry out cooperative work in
virtual machines with different specifications to implement case of emergency has not been well solved. In this paper, we
data processing for various terminals [8]. Meanwhile, ECSs first construct the ECS network architecture utilizing the SDN
can also form a collaborative network to achieve information paradigm to achieve the global coverage, especially in remote
sharing and unified cross-regional terminal scheduling [9]. areas (e.g., desert, ocean, forest, etc.), where a terrestrial net-
However, since LEO satellites are moving at a relatively high work is impossible or impractical to reach. The GEO satellites
speed, its topology and coverage area in the next time slot may with wide coverage are used as the SDN controller to achieve
change, leading to the reconfiguration of the resources in real-time control of the LEO constellation. The edge servers
ECSs. The resources and time spent on reconfiguration may deployed in ECSs are mainly responsible for data processing
impact the ability of ECS for real-time data processing. The of intelligent terminals in the coverage region. When terminals
routing strategies of resource scheduling and information accessing, the clustering algorithm carried by ECS can classi-
sharing for ECS network are also problems when facing emer- fy resource requirements of various terminals to help achieve a
gency [10]. Therefore, a dynamic resource scheduling scheme precise division of computing resources. The division config-
in ECS network is needed to help the ECS realize real-time urations actually have similarity in adjacent regions. So the
data processing for intelligent terminals. configuration, as a template, can be transferred to other nodes
How to carry out fast resource allocation for each accessing of the fleet by SDN controller to reduce the resource config-
terminal is the main challenge. The scheduling of resources in uration time of the entire ECS network. Moreover, some re-
satellite networks is an ongoing research area and there have sources of ECS are reserved for emergency. The ECS collab-
been some important studies. Authors in [11] considered the oration for multi-region terminal scheduling can be realized
cooperative mechanism of relay satellites deployed in GEO with the global knowledge of GEO controller, when faced
and LEO and proposed a multiple access and resource alloca- with large-scale emergencies, such as storms, tsunamis, and
tion strategy for GEO relay in LEO satellite network. But they other natural disasters. Actually, the resource scheduling
didn’t consider the high data transmission latency of GEO schemes both ECS-ECS and ECS-terminals are comprehen-
satellites, about 500 ms round trip time, which could affect sively considered. In conclusion, this fine-grained network
the real-time management for terminals. Kodheli et al. [12] slicing scheme can allocate resources reasonably for accessing
proposed a resource allocation approach over LEO satellite terminals and reduce resource waste.
architecture in order to reduce the high values of Doppler shift The rest of this paper is organized as follows. Section 2
of terminals. Although the performance degradation of LEO introduces the edge computing satellite network architecture
constellation was reduced, the dynamic resource scheduling based on SDN framework. Section 3 will construct the math-
of LEO constellation was not solved to some extent. For ematical model of resource allocation in ECS and propose the
multi-ECS collaboration, Sinha et al. [13] presented a multi- ECS collaboration routing strategy. The simulation results and
agent-based modeling of LEO satellite network. Satellites analysis are shown in section 4. We then conclude our work in
were modeled as autonomous agents and could collaborate Section 5.
with other satellite agents. The allocation of tasks by the
agents was modeled as a distributed constraint optimization
problem. But self-organizing networks may affect other satel- 2 Edge computing satellite network
lites, causing unnecessary link congestion. With the develop- architecture
ment of information and communication technologies, the
idea of software-defined scheduling strategy for LEO constel- In this section, we will introduce the edge computing satellite
lation has obtained more attention. The authors in [14] pro- network architecture to exhibit data interaction in different
posed a software-defined model to manage and orchestrate layers and the framework of mobile edge computing servers
networking, caching, and computing resources jointly in in ECS, as shown in Fig. 1. The right half of the figure ex-
LEO network. They formulated the joint resources allocation presses a three-layer space-air-ground integrated network,
problem as a joint optimization problem and used a deep Q- which are GEO satellite layer, LEO satellite layer, and air-
learning approach to solve it. Sheng et al. [15] constructed a ground layer.
novel graph model to describe the evolution of multi-
dimensional resources in a satellite network. They proposed Air-ground layer Intelligent terminals are installed on electron-
an optimal resource allocation strategy to facilitate efficient ic equipment, such as vehicles, UAVs, ships and sea sensors,
cooperation among various resources. to provide various intelligent services. The air-ground layer is
It can be seen from the above that using software-defined the main accommodate layer of these terminals. Nowadays
networking (SDN) architecture to realize dynamic allocation intelligent terminals have been distributed in every corner of
Mobile Netw Appl

Fig. 1 The edge computing Delay Bandwidth Connection Emergency GEO satellite plane
Application
satellite network architecture service demand demand time demand control
VM VM VM VM

Edge
GEO SDN controller
computing Delay Bandwidth Connection Service
application function fucntion time function management
management
system virtualization and resource allocation
Edge computing application platform ECS

LEO satellite plane

Edge Edge computing virtual resource
computing Air-ground plane
infrastructure
management SDN Control
Edge computing hardware resource
system
Edge computing hosting infrastructure Edge Computing Control

the ground-air layer, mainly responsible for communication, GEO satellite layer Considering the high dynamics of LEO
data collection, and data processing. However, new technolo- satellites, it is difficult to load and update services of ECS with
gies, such as augmented reality (AR), virtual reality (VR) and a ground station. Meanwhile, the connection of ISLs needs
artificial intelligence (AI), are emerging. The types and vol- global knowledge of LEO constellation. Therefore, the
umes of data collected by terminals are growing rapidly. The software-defined network (SDN) paradigm is established in
demand for terminal computing power is growing, which is the two-layer satellite network. SDN architecture decouples
contradictory to the smaller volume of terminals. To solve this the network control and forwarding functions, enabling net-
problem, the land terminal can choose to transfer the comput- work control to become programmable. Considering the dif-
ing unit to the nearby mobile edge computing (MEC) servers ferent characteristics of satellites, the GEO satellites with wide
to realize data processing with low latency. But in other envi- coverage are used as the SDN controller to achieve real-time
ronments, such as deserts, oceans, and air, MEC servers are control of the LEO constellation using the OpenFlow proto-
difficult to deploy. With the development of LEO satellite col. The controller can hold the topology of the global LEO
technology, it is increasingly popular to deploy MEC server constellation while maintaining constant contact with the
on LEO satellite to realize data processing and terminal con- ground data center. In this way, the updating patches of LEO
trol in harsh environments. satellites can be timely sent to LEO satellites through the GEO
controller. More importantly, when an emergency happens
LEO satellite layer LEO satellite network communication such as earthquake and tsunami, GEO satellites can receive
technology has been viewed as one of the most enabling the solutions uploaded by the ground command center quickly
technologies for future communication, which have some and stably. With the advantage of full coverage, GEO satellites
advantages that the ground communication network does can broadcast different scheduling schemes to relevant LEO
not have. In LEO satellite layer, the LEO constellation satellites to achieve precise direction in real time.
achieves the global coverage of earth to provide communi-
cation service, especially in the remote land and the ocean Edge computing satellite The edge servers deployed in LEO
where the ground base station cannot build. In order to realize satellites can form ECS network to improve the management
real-time data processing and terminal control under severe for terminals in the region. Technologies of KVM and Open
natural conditions, the edge computing paradigm is lever- vswitch are applied in the programmable server to help ECS
aged to LEO satellites to form edge computing satellite achieve efficient resource management and division. The left
(ECS). Utilizing the computing power of edge servers, ter- half of Fig. 1 expresses the possible infrastructure and appli-
minal data can be analyzed and fed back directly on ECS in cation of ECS. The infrastructure platform mainly realizes the
real time, without transmission to the ground data center. virtualization of computing hardware resources to provide
This has significant advantages for intelligent terminals de- more flexible computing services. The application platform
ployed on UAVs, ships, and sensors in the no man’s land. integrates the computing resources of ECS to realize various
Further, edge servers are programmable and can be manually computing services for accessing terminals. Especially, the
developed and updated various application services to cope ECS analyze different computing resource demands of termi-
with different types of terminals. ECSs can share data with nals and carry out corresponding resource allocation functions
each other through inter-satellite links (ISLs) to realize co- to realize the management of terminals, such as delay func-
operative control of terminals, such as UAV cluster. tion, bandwidth function and connection time function. The
 Mobile Netw Appl

specific resource division scheme is proposed in section 3. sensitivity, bandwidth sensitivity and connection time sensi-
Moreover, a certain part of resources is reserved in ECS for tivity. The ECS uses a clustering algorithm, such as the K-
emergencies. When wide-range emergency, such as earth- means algorithm, to obtain the number and types of different
quake and tsunami, happens, the ECS receives the strategy terminals in the region to guide the resource allocation of
from the GEO controller and generates the emergency control VMs.
VM utilizing reserved resources to perform the final solution
of terminal scheduling. In a certain extent, the resource Delay sensitivity Without considering the time taken by the
virtualization and programmable scheduling functions in signal transmission and reception, the time taken by transmit-
ECS guarantee flexible computing services for terminals. ting the information from the terminal to the ECS is assumed
The edge computing satellite network architecture uses as the transmission delay diE(t):
ECSs to help terminals to process data on the edge of the qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
network with low latency. Meanwhile, the SDN framework ðxE ðt Þ þ xi ðt ÞÞ2 þ ðyE ðt Þ−yi ðt ÞÞ2 þ ðzE ðt ÞÞ2
is leveraged for solution allocation of emergency. The whole d iE ðt Þ ¼ ð1Þ
c
architecture can realize the intelligent management of termi-
nals around the world. where (xi(t), yi(t), zi(t)) and (xE(t), yE(t), zE(t)) respectively rep-
resent the three-dimensional coordinates of ECS E and termi-
nal i, and they are all functions of time. c is the speed of light.
3 Mathematical model It can be seen from the above equation that the delay is lower
when terminal i runs at a high altitude. If terminal i has a fast
In edge computing satellite network, each ECS can connect speed it will have a large value of diE'(t)
multiple terminals by the user data links (UDLs). The ECS
0 Δd iE ðt Þ
generates the corresponding VM to provide computing ser- d iE ðt Þ ¼ ð2Þ
Δt
vices for the terminal. Therefore, a question needs to be con-
0
sidered that how to allocate the computing resources of ECSs The value of d iE ðt Þ is denoted by DiE to express the delay
for the terminals, so as to not only meet the computing de- 0
sensitivity of terminal i. Based on d iE ðt Þ, the relative position
mand but also realize reasonable resource configuration.
and motion direction of terminals and satellites can be predict-
What’s more, various types of terminal data will be transmit-
ed, helping to plan more appropriate ISLs. The delay sensitiv-
ted to the ground data center and help it to monitor the envi-
ity of n terminals can be expressed as {DiE, D2E, ...DnE). For
ronmental conditions in real time. When emergencies occur
convenient comparison with the other indicators, we normal-
(such as storm, tsunami), the terrestrial data center can make
ize the DiE as D*iE :
decisions in time and utilize GEO controller to send corre-
sponding responses to each ECS. Then the ECS inform in- DiE −Dmin
D*iE ¼  100 ð3Þ
volved terminals to change the route to realize intelligent Dmax −Dmin
scheduling. Therefore, efficient ISL between ECSs is required
to establish the data-sharing network to ensure data transmis- where Dmax and Dmin are the maximum and minimum values
sion. This section we first discuss dynamic resource allocation in the dataset. In order to obtain higher data accuracy, the
strategy in ECS for various terminal access, considering dif- normalized value is magnified by 100 times. It can be known
ferent sensitivity of terminals in delay, bandwidth and connec- form above that terminals such as UAVs and airplanes may
tion time. Then, an ISL establishment scheme is proposed to have higher values of D*iE , considering their high-speed rela-
achieve efficient data transmission in emergency situations. tive motion with the ECS. Therefore, these terminals had bet-
ter be grouped together for fast data calculation and transmis-
3.1 Resource allocation in the edge control satellite sion to reduce the influence of signal distortion.

Different terminals have different data computing demands. Bandwidth sensitivity Terminals such as data centers need to
For example, due to the relatively high-speed motion between collect all kinds of information from remote terminals, so the
UAVs and satellites, low delay data transmission is required to ECS connected with these terminals is required to provide
reduce signal distortion. Considering the data center has a higher transmission bandwidth. Such terminals can be
fixed location and need to receive various types of terminal expressed as having high bandwidth sensitivity. The data
data, there is a high bandwidth demand for it when accessing transmission rate always depends on the bandwidth provided
ECS. Frequent link switching may result in partial data loss, by the ECSs’ transceiver and the unit is Mbps. For example,
so for sensor nodes and ships, it is necessary to select ECS for the total bandwidth BEtotalof ECS E is 20Mbps, and 5Mbps
data processing with long connection time. Therefore, intelli- has been used to forward other tasks. Meanwhile, the trans-
gent terminals can be divided into three types: delay ceiver capacity B itotal of terminal i is 18Mbps. If the
Mobile Netw Appl

communication link is established between E and i, the max-

imum available bandwidth BEi is 15Mbps, which is the small- T iE −T min
er available Bvalue for both nodes: T *iE ¼  100 ð7Þ
T max −T min
BEi ¼ minðBE ; Bi Þ ð4Þ
where Tmax and Tmin are the maximum and minimum
The value of BEican be used to describe the bandwidth
values in the dataset. Finally, the terminal indicator set
sensitivity of terminal i. The bandwidth sensitivity of n
can be expressed as X:
terminals can be expressed as {B 1E , B 2E , ...B nE }. The
higher the value, the stronger the bandwidth sensitivity. X ¼ fx 1 ; x 2 ; x n g ð8Þ
For convenient comparison with the other indicators, we  * * *
normalize the BiE as B*iE : where xi ¼ DiE ; BiE ; T iE is a three-dimension vector to
represent the sensitivity characteristics of terminal i. In
BiE −Bmin order to achieve the terminal clustering, Disij is defined
B*iE ¼  100 ð5Þ as the Euclid distance between terminals of the same
Bmax −Bmin
cluster:
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
     
2 2 2
where Bmax and Bmin are the maximum and minimum Dis
ij ¼ D*iE −D*jE þ B*iE −B*jE þ T *iE −T *jE ð9Þ
values in the dataset. In general, the ECS will expand
the resources and computing power of corresponding
Suppose the classified cluster is {C1, C2, ...Cn}, the average
VMs to realize the fast transmission of terminal data when
distance Dis in the same cluster can be expressed as
flying over the terminal with high bandwidth sensitivity.
∑ni ¼1; j≠i Disij
Connection time sensitivity Frequent link switching may Disij ¼ ð10Þ
cin
result in partial data loss of terminals like sensors and
ships, so these terminals have high connection time sen- According to the above equations, the standard deviation
sitivity. The link connection time starts with the estab- SDi of the terminals in the same cluster is:
lishment of transmission links and ends with the satel- vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
lites disappeared, where the disappearance means that u  2
u n n
two nodes have lost their physical visibility. We set Re t ∑i 1 ∑ j 1; j≠i Disij −Disij
SDi ¼ ð11Þ
as the radius of the Earth, E as the ECS flying around n
the Earth, i as the terminal and L as the distance of E
and i. When the communication link between E and i is In the cluster i, the smaller the SDi value, the higher
tangent to Re, the critical value is reached. The value of the similarity of the terminals. So SDi is utilized to
L at this time reaches the maximum value Lmax. The evaluate the clustering results. In the results obtained
next moment the ECS is not physically visible and it by the clustering algorithm, the minimum of the maxi-
cannot maintain the transmission links. We define the mum standard deviation of all clusters is pursued:
critical time when link establishment as T0 and when
two nodes lose physical visibility as Tmax, the link con- minfmaxSDi g; i ¼ 1; 2; 3; ::: ð12Þ
nection time TEi between ECS E and terminal i can be
expressed as:
Considering that resources in edge servers should be
used for terminal data calculation as far as possible, the
T Ei ¼ T max −T 0 ð6Þ clustering algorithm will choose the K-means algorithm,
which has simple and fast clustering ability. At the same
In fact, the value of is related to the velocity of the time, the number of initial clusters and clustering centers
terminals. UAVs and airplanes with high-speed move- can be preset to reduce the computation complexity and
ment have lower Tmaxvalues, while terminals such as clustering error. In this three-dimensional case, the qua-
ships and buoy sensors have higher Tmax values because dratic mean of each dimensional data is used to generate
of the low relative motion. The connection time sensi- the initial cluster center:
tivity of n terminals can be expressed as {T1E, T2E, TnE}.
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
The higher the value, the stronger the connection time 2 2 2
D*1E þ D*2E þ ⋯ þ D*nE
sensitivity. For convenient comparison with the other Dsq ¼ ð13Þ
indicators, we normalize TiE as T *iE :
 Mobile Netw Appl

scheme aims to reduce the data queuing time when managing

massive terminals in different types. Precise on-demand dis-
tribution reduces the resource waste. Moreover, reasonable
use of resources can also help ECS to realize the other appli-
cation services.

3.2 Dynamic adjustment

The above scheme helps the ECS to realize the reasonable

division of resources. However, terminals in different regions
may have different characteristics. For example, terrestrial ter-
minals like the data center use the ECS to collect data from
remote terminals but ships in the ocean may utilize the ECS to
guide the route. These will lead to different resource division
configurations of one ECS when flying over different regions.
It will consume massive computing resources if the ECS run
clustering algorithm in every time slot. The initial access of
terminals will also be affected. Therefore, a reference config-
uration transfer scheme is proposed to reduce the unnecessary
resource consumption of ECSs. As Fig. 2 shows, in two-layer
satellite network, the GEO satellite, as the SDN controller, can
acquire the global knowledge of the network topology and
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi information of ECSs. Meanwhile, the trajectory of each ECS
2 2 is fixed, which leads to the periodic change of the whole
B1E þ B*2E þ ⋯ þ B*nE
2 *
Bsq ¼ ð14Þ satellite network topology. So the ECS also changes periodi-
n
cally for a certain region of the earth. Suppose that the ECS
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi fleet is X = (x1, x2, ...xm) and the control time of xi is ti:
2 2 2
T *1E þ T *2E þ ⋯T *nE
T sq ¼ ð15Þ x1 ¼ t 1 ; x2 ¼ t 2 ; x3 ¼ t 3 ; :::; xm ¼ t m ð16Þ
n

The whole control period Pe is

The initial coordinates are(Dsq, 0, 0). , . (0, Bsq, 0) and (0,
0, Tsq). In this paper, the advanced K-means algorithm (AKG) Pe ¼ ∑m
i¼1 t i ð17Þ
is used to cluster the terminals in the following steps:
It means satellite xi will be the ECS of the region once again
Step 1. Select(Dsq, 0, 0), (0, Bsq, 0)and(0, 0, Tsq) as the cen- after Pe time. This periodic movement can be used to ensure
ters of three clusters; the continuous control of terminals. Further, the last ECS can
Step 2. Calculate the distance from the remaining terminals provide its configuration to the next ECS as reference.
{x1, x2, ...xn}to each cluster center and classify these Utilizing the GEO controller, the current resource division
terminals into the nearest cluster center; configuration is transferred at the end of ti to the next ECS
Step 3. Calculate the average values Disij of terminals in for initial allocation of resources. There is no need for the next
each cluster respectively and designate them as the
new cluster centers.
Step 4. Iterate Step 2 and Step 3 until reaching the threshold
or cluster centers are not changed.

The terminal classification algorithm are shown as

Algorithm 1. Through the AKG algorithm, terminals are di-
vided into three categories as delay sensitivity, bandwidth
sensitivity and connection time sensitivity. The ECS calcu-
lates resource requirements of each type of terminals for VM
resource allocation, so that each terminal can obtain appropri-
ate computing resources. This dynamic resource allocation Fig. 2 ECS configuration deliver model
Mobile Netw Appl

ECS to collect terminal information again and then carry out quickly form ISLs between specific ECSs for data transmis-
the AKG algorithm to allocate resources, but only need to sion. Due to the complex topology of the LEO constellation, it
make small adjustments on the basis of the obtained configu- is necessary for the GEO controller to make the incidence
ration. This dynamic adjustment scheme greatly reduces the matrix (IM) of the current ECS network. First, all the ECSs
recalculation resources of ECS. The continuous control of are numbered in the constellation. The orbit number is set as i
terminals can be achieved directly according to the dynamic and the satellite number is set as j. For all ECSs, there are
adjustment.

8 9
> E ; E 12 ; :::; E1 j >
3.3 ECS collaboration < 11 =
E21 ;
IM ¼ E 22 ; :::; E2 j ð18Þ
>
: ::: >
;
Normally, each ECS is responsible for the management of E i1 ; Ei2 ; :::; E ij
terminals in a certain region. The terminal transfers the com-
puting unit to ECS and the ECS uses the data processing
ability to realize regional control. However, in the face of
large-scale emergencies, such as storms, tsunamis, and other
natural disasters, the limited computing resources of one sin-
gle ECS may not be able to calculate an effective response in a
short time for massive terminal scheduling. Moreover, even if
one ECS implements the scheduling of terminals in the region,
its scheme may conflict with other ECSs, resulting in conges-
tion and unexpected losses. Therefore, when large-scale emer-
gencies happen, it is necessary to carry out unified scheduling
for the terminals that may be affected to ensure a fast and
efficient response. Using the collaboration of ECSs, the pro-
posed architecture has a potential opportunity to change this
impasse.
The terrestrial data center have powerful processing capa-
bilities for massive data and can obtain computational results
in milliseconds. In the proposed architecture, various terminal
data from different regions are uploaded to ECSs and then
transmitted to the data center via inter-satellite links (ISLs),
as shown in Fig. 3. Real-time data processing and analysis
take place in the data center. When an emergency occurs, the
data center plans the responses immediately and sends differ-
ent emergency responses to corresponding ECSs. After ECS
receiving, it starts the emergency control VM to further anal-
ysis and send the warning information and scheduling plan to
each terminal in the management region.
The above scheme aims to utilize ECS cooperation for where Eij represents the jth satellite in the ith orbit. After the
rapid emergency response. The core of the strategy is how to ECSs to be connected are determined, it is assumed that the
ECS connected with the data center is Eij. If its target ECS are
T1 A1 A2 A3 far away, Eij needs to determine some nearby relay satellites
Accessing terminals: {Li,j, } to forward the scheme. Here, we consider the breadth-
A6 A5 A4 first-search-based spanning tree algorithm (BFST) for the es-
B1 B2 B3 tablishment of ISLs. The BFST method can not only establish
Edge control satellites:
the shortest ISLs from the source ECS to the destination ECS,
B6 B5 B4 T2 but also have some practical potential that other similar algo-
Relay satellites: rithms do not have. First is that the ISLs are not multiplexed.
C1 C2 C3 Algorithms like Dijkstra, gradient boosting decision tree
(GBDT) and depth-first-search-based spanning tree (DFST)
T3 C6 C5 C4 focus on building one shortest link that connects all nodes in
Fig. 3 ECS collaboration model series. Though the transmission delay is small, each ECS
 Mobile Netw Appl

receives a lot of unnecessary data, which belongs to other Table 1 Orbital parameters of satellites
ECSs. Meanwhile, the time of data received by each ECS is Type of satellite Orbit height/km Number Orbit inclination
also different, which will cause the asynchronous response
and affect the correct implementation of the scheme. So the GEO 36,000 4 0
ISLs had better not be multiplexed to achieve synchronize LEO 895.5 6 × 11 90
response of each ECS. The next is that BFST has enough
alternative ISLs with high quality. In a single connecting link,
the rebuild of ISL will cause higher delay increase if one ECS  
H Eij ; E i ; j ; > Re ð19Þ
in the link is broken. So the quality of alternative links can be
used as an indicator when ISL building. Taking into account where H(Lij, Li,j, ) express the distance between the Geocenter
the above analysis, some link selection conditions can be con- and (Eij, Ei, j, ). In the LEO satellite constellation, considering
sidered to reduce the computation of BFST when carrying out the real-time performance of data transmission, we choose the
the ISL establishment algorithm. distance between satellites as the edge weight in BFST:
According to the visibility theory described above, it is    
necessary to determine whether the selected satellite can com- w E ij ; Ei ; j ; ¼ D E ij ; Ei ; j ; ð20Þ
municate with each other. That is to determine if(Eij, Ei,j, )
exist. The judgment formula is shown as follows: whereD(Eij, Ei,j, )is the distance of (Eij, Ei,j, ). Considering the

Fig. 4 The flow chart of the

dynamic resource scheduling
scheme
Mobile Netw Appl

Fig. 5 Two-layer satellite simulation model

BFST algorithm can establish multiple branches to transmit

Fig. 7 Comparison of clustering results and original data
data together, each branch is hoped to transmit data in the
shortest time, especially when an emergency happens.
Therefore, the final algorithm can be modeled as follows:
4 Simulation Results and Analysis
m n m n   The simulation is carried out over a predictable two-layer
min ∑ ∑ ∑ ∑ w E ij ; E i0 j0
i¼1 j¼1 i0 ¼1 j0 ¼1 satellite network with the use of the Satellite Tool Kit (STK)
 
s:t: H E ij ; E i0 j0 > Re simulator. The satellite model is established with reference to
b1 ≤ i ≤b2 ð21Þ iridium constellation to guarantee the practical application.
b3 ≤ j≤ b4 The satellite and orbital parameters are shown in Table 1 and
0
 ≤i ≤i þ
i−1 1    the scene generated in STK is shown in Fig. 5. Meanwhile, we
D Lij ; Li0 j0 ≤ min D Lij ; Li2; j2 randomly created 100 UAV terminal nodes, 200 ship nodes
and 200 sensor nodes in the Pacific Ocean using STK. Then
where {bi} are the borders of the satellite network with we connect STK with MATLAB to obtain the satellites and
minimal inclusion of ECSs to be connected. Algorithm 2 rep- terminal data, where the calculation and comparison of the
resents the BFST inter-satellite link establishment algorithm proposed scheme and other methods are also carried out. In
steps, and the detail process from resource allocation to ISL order to prevent overfitting, the initial data we set is irregular
building of ECS is shown in Fig. 4. to ensure diversity. Meanwhile, in the simulation experiments,
random noise is added to the data to ensure the randomness.

Fig. 6 Error comparison of two clustering algorithm Fig. 8 Comparison of the maximum ISL hops
 Mobile Netw Appl

Each terminal node is assigned different values of three sen-

sitivities according to their characteristics of resource demand.
In the simulation, we first verified the classification preci-
sion of the proposed advanced k-means algorithm (AKG). As
shown in Fig. 6, we compare the clustering errors of the basic
k-means algorithm and the proposed AKG algorithm with the
increase of terminal numbers. The trends of error change are
plotted by curves. It can be seen that the clustering error of
AKG is significantly lower than that of basic k- means algo-
rithm. The error is almost always half of the basic k-means
method. This is because the clustering center at the initializa-
tion of the AKG has been calculated regularly, while the basic
k-means algorithm just randomly selects the clustering center.
The selection of clustering center will greatly affect the accu-
racy and computational complexity of the algorithm.
Therefore, by considering the characteristics of three data
types, we directly calculate the initial clustering center, so as Fig. 10 Comparison of the available bandwidth
to improve the classification accuracy of the algorithm. The
clustering result is expressed as Fig. 7, which shows a good with the increase of the involved ECSs. The growth of BFST
terminal classification result. When the emergency happens, is the slowest, followed by TSA and SPDA method. This is
reliable communication links need to be established between because the BFST tries to avoid the reuse of links when path
the ground centers and oceangoing terminals through ECSs to finding, so that each ISL of ECS maintains a similar number
transmit the decision. The quality of the ISLs determines de- of hops. The TSA method firstly finds a relay satellite node
cision transmission. Next, we will evaluate the proposed and then connects all involved ECSs to the relay node to
BFST scheme in terms of link hops, delay, bandwidth, and realize ISL establishment. Although the ECS interconnection
robustness, compared with current mainstream ISL construc- is realized, the link between the involved ECS and the ground
tion schemes including the shortest path Dijkstra algorithm center becomes more complex, resulting in more link hops.
(SPDA) and the traveling salesman algorithm (TSA). The SPDA is to establish the shortest path between all ECSs.
As information is transmitted through ISLs, a large number But the link between the initial ECS and the end ECS is too
of link hops mean that the information needs to be forwarded long, which will greatly affect the data transmission of the end
by multiple ECS before reaching the target satellite. The in- node. Overall, the proposed BFST method has relatively good
crease in satellite hops may result in more packet loss. So the data integrity with minimum ISL hops.
maximum ISL hops is an indicator of the ISLs quality. As Fig. Transmission delay has a direct influence on real-time de-
8 shows, the maximum ISL hops of all three methods grow liver of decision. Figure 9 compares the maximum transmis-
sion delay of the three methods. It can be seen that the BFST

Fig. 9 Comparison of the maximum transmission delay Fig. 11 Comparison of link stability
Mobile Netw Appl

has the minimum transmission delay, followed by TSA, and both methods have established a long single connected
finally SPDA method, which is similar to the change of ISL link. It can be known that a network of single connected
hops number. But SPDA has much higher link delay than the links is the most vulnerable. If any sublink malfunctions,
other two methods. This is because both BFST and TSA the entire network is down. Different from the SPDA meth-
methods choose to use multiple paths to transmit data. The od, the BFST avoids the connection between ECS nodes as
increase of branch will help the main path to amortize most of much as possible, so that all data is amortized in each link
transmitting data, thus reducing the transmission delay. In instead of being transmitted together, achieving better ECS
BFST, each ECS corresponds to a branch, which minimizes network stability. It can be seen that the link stability of
transmission delay. multi-branch link planning scheme is always better than
The ISL available bandwidth reflects the data transfer that of single-link scheme. The robustness represents the
rate. In the simulation, the bandwidth values of each performance of continuous service of network and can be
available link were randomly assigned between 10- shown by the ISL connection time. Figure 12 expresses the
50Mbps. As Fig. 10 shows, with the increase of involved change of ISL connection time for three methods as the
ECS, the ISL available bandwidth of all three methods number of involved ECS increases. It can be seen that the
decreases. This is because the increase of ECS creates proposed BFST method always maintains more link con-
more links, while the ISL available bandwidth is the min- nection time, compared with the other two methods. As the
imum of all link bandwidth. Since SPDA method can be SPDA is a single connected link, its ISL connection time
seen as the superposition of multiple simple links, its total decreases with the growth of the ECS nodes. In order to
available bandwidth decreases the fastest. The BFST aims maintain less total link delay, the TSA selection of relay
to find a non-reusable ISL for each ECS, so that the de- node is too close to the involved ECS region, resulting in
cision for each ECS can be transmitted in the exclusive less connection time of link between ECS in the ground
link to achieve the maximum utilization of bandwidth. region and relay node. The proposed BFST method is a
In addition to considering the performance of ISL links, combination of multiple similar sublinks. The connection
the stability and robustness of ECS networks constructed time of each sublink has little difference, bring slowly
in emergency situations are also important indicators to curve drop.
evaluate the efficiency of data transmission and sharing. All the above evaluation show that the proposed AKG
Stability refers to the ability of the ECS network to persist algorithm provides higher clustering precision and the BFST
interconnection over time. The evaluation of stability can scheme in ISL deployment can not only realize data transmis-
be expressed by the number of link reconfigurations over sion with low latency and high bandwidth but also achieve
time. As shown in Fig. 11, the number of link reconstruc- stability and robustness of ECS networks constructed in emer-
tion of the three methods increase with time change. At gency situations, compared with other two methods.
each time slot, the SPDA method has the highest number
of link reconstruction, followed by TSA, and BFST is the
least. At the early stage, the number of link reconstruction
of SPDA and TSA methods rise rapidly. This is because
5 Conclusions

This paper studies the resource scheduling problem in the

edge computing satellite network. In contrast to the previous
study, we jointly consider resource division strategy in edge
computing satellite (ECS) and the ECS collaborative network
establishment in emergency situations. We regard different
sensitivity of terminals in delay, bandwidth and connection
time as resource allocation factors and propose an advanced
K-means algorithm (AKG) to guide the resource division in
ECS. Meanwhile, we improve the algorithm of breadth-first-
search-based spanning tree (BFST) for the establishment of
ISLs to realize the dynamic scheduling of information and
resources in emergency situations. The simulation results
show that the proposed dynamic resource scheduling scheme
is promising.

Acknowledgments This work was supported in part by the National

Natural Science Foundation of China (No. 61571104), the Sichuan
Fig. 12 Comparison of link robustness Science and Technology Program (No. 2018JY0539), the Key projects
 Mobile Netw Appl

of the Sichuan Provincial Education Department (No.18ZA0219), the 7. Ren J, Pan Y, Goscinski A et al (2018) Edge computing for the
Fundamental Research Funds for the Central Universities (No. internet of things. IEEE Netw 32(1):6–7
ZYGX2017KYQD170), and the Innovation Funding (No. 8. Li H, Ota K, Dong M (2018) Learning IoT in edge: deep learning
2018510007000134). The authors wish to thank the reviewers for their for the internet of things with edge computing. IEEE Netw 32(1):
helpful comments. Dr. Dingde Jiang is corresponding author of this paper 96–101
(email: merry_99@sina.com). 9. Chen X, Shi Q, Yang L et al (2018) ThriftyEdge: resource-efficient
edge computing for intelligent IoT applications. IEEE Netw 32(1):
61–65
10. Tran TX, Pompili D (2019) Joint task offloading and resource allo-
References cation for multi-server mobile-edge computing networks. IEEE
Trans Veh Technol 68(1):856–868
1. Qu Z, Zhang G, Cao H et al (2017) LEO satellite constellation for 11. Du J, Jiang C, Wang J et al (2017) Resource allocation in space
internet of things. IEEE Access 5:18391–18401 multiaccess systems. IEEE Trans Aerosp Electron Syst 53(2):598–
2. Liu J, Shi Y, Zhao L et al (2018) Joint placement of controllers and 618
gateways in SDN-enabled 5G-satellite integrated network. IEEE J 12. Kodheli O, Andrenacci S, Maturo N et al (2018) Resource alloca-
Sel Area Comm 36(2):221–232 tion approach for differential doppler reduction in NB-IoT over
3. Zhang Z, Jiang C, Guo S et al (2018) Temporal centrality-balanced LEO satellite. In: Proc. ASMSSPSC'18, pp 1–8
traffic management for space satellite networks. IEEE Trans Veh 13. Sinha PK, Dutta A (2016) Multi-satellite task allocation algorithm
Technol 67(5):4427–4439 for earth observation, in Proc. TENCON'16, pp 403–408
4. Wang P, Zhang X, Zhang S et al (2019) Time-expanded graph 14. Qiu C, Yao H, Yu R et al (2019) Deep q-learning aided networking,
based resource allocation over the satellite networks. IEEE Wirel caching, and computing resources allocation in software-defined
Commun Le 8(2):360–363 satellite-terrestrial networks. IEEE Trans Veh Technol
5. Tang F, Zhang H, Fu L et al (2019) Multipath cooperative routing 15. Sheng M, Wang Y, Li J et al (2017) Toward a flexible and
with efficient acknowledgement for LEO satellite networks. IEEE reconfigurable broadband satellite network: resource management
Trans Mob Comput 8(1):179–192 architecture and strategies. IEEE Wirel Commun 24(4):127–133
6. Tran TX, Hajisami A, Pandey P et al (2017) Collaborative mobile
edge computing in 5G networks: new paradigms, scenarios, and Publisher’s note Springer Nature remains neutral with regard to jurisdic-
challenges. IEEE Commun Mag 55(4):54–61 tional claims in published maps and institutional affiliations.