IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. 18, NO.

2, MARCH/APRIL 2025 839

Joint Service Deployment and Task Offloading for

Datacenters With Edge Heterogeneous Servers
Fu Xiao , Senior Member, IEEE, Weibei Fan , Member, IEEE, Lei Han , Tie Qiu , Senior Member, IEEE,
and Xiuzhen Cheng , Fellow, IEEE

Abstract—Mobile edge computing (MEC) can improve execution offloaded to the data center at the edge of the network for mobile
efficiency and reduce overhead for offloading computing tasks to edge computing (MEC) [2], [3]. Compared with traditional
edge servers with more resources. In the microservice system, the cloud data centers (DCs), it is geographically closer to users
current research only considers the cross segment communication
cost of computing tasks, does not consider the case of the same and terminal devices, which can reduce the delay of application
end, and ignores the discovery and invocation optimization of service requests and bring higher service quality to users.
associated services. In this paper, we propose CACO, which is a In the current research on application division based on mi-
novel content-aware classification offloading framework for MEC croservice architecture, most of them are considered in the cloud
based on correlation matrix. CACO first designs an adaptive service computing environment [10], [11]. Firstly, they are all executed
discovery model, which can make timely response and adjustment
to the changes of the external environment. It then investigates an inside cloud centers when microservices are divided into differ-
efficient affinity matrix based service discovery algorithm, which ent partitions. Compared with the scheduling in edge computing,
expresses the association relationship between services by con- the time loss of data transmission and communication between
structing a service association matrix. In addition, CACO constructs microservices during execution is often ignored. Secondly, in the
a relational model by giving different weight coefficients to the previous studies based on edge computing, most of them are still
delay and energy loss, which improves the delay and energy loss
of message processing in a satisfying manner. Simulation results based on coarse-grained offloading [4]. In the target selection of
indicate that CACO reduces the total traffic of redundant messages task offloading, only the cloud center and local edge devices are
by 46.2% ∼76.5% , respectively compared with state-of-the-art often considered. The coarse-grained offloading method is also
solutions. Testbed benchmarks show that it can also improve the not applicable to the new edge computing scenario where a large
stability by reducing control overhead by 34.5% ∼81.6% . number of lightweight computing nodes are introduced [7], [8].
Index Terms—Mobile edge computing, service deployment, task In this paper, we investigate a new classification problem that
offloading, performance evaluation. prioritizes secure messages to maximize latency and energy loss,
while ensuring user experience and acceptable time complexity.
Three major challenges must be addressed. The first challenge is
I. INTRODUCTION
to design an adaptive service composition model driven by QoS.
ITH the development of edge computing technology, a
W growing number of terminal device applications can be
The previous work [13] and [14] abstracted the QoS expecta-
tion of service composition as the objective function of mixed
integer linear programming (MILP). The QoS attributes of Mi-
Received 30 September 2024; revised 25 December 2024; accepted 1 February croservices and composite service are abstracted as constraints,
2025. Date of publication 5 February 2025; date of current version 10 April and the optimal solution of the objective function is obtained
2025. This work was supported in part by the Major Scientific Instruments within the scope of constraints. However, this assumption may
and Equipments Development Project of National Natural Science Foundation
of China under Grant 62427809, in part by the National Key Research and lead to biased choices when applied to our situation. Most
Development Program of China under Grant 2024YFB2906700, in part by the QoS adaptive service portfolios [15] provide optimal composite
National Science Fund for Distinguished Young Scholars under Grant 62125203, services, rather than services that are closest to user service
in part by the Key Program of Natural Science Foundation of Jiangsu under Grant
BK20243053 and Grant 24KJA520006, and in part by the Natural Science Foun- level agreements. Therefore, this method is particularly complex
dation of China under Grant 62102193 and Grant 62372248. (Corresponding when there are a large number of services and QoS attributes.
author: Weibei Fan.) The second challenge is to develop a reasonable offloading
Fu Xiao, Weibei Fan, and Lei Han are with the College of Computer, Nanjing
University of Posts and Telecommunications, Nanjing 210023, China, and also strategy that balances energy consumption and latency. A tradi-
with the Shanghai Key Laboratory of Trustworthy Computing, East China tional solution is energy optimized offloading, which is superior
Normal University, Shanghai 200050, China (e-mail: [email protected]; wb- to local computing and complete offloading methods in terms of
[email protected]; [email protected]).
Tie Qiu is with the School of Computer Science and Technology, College of energy consumption and latency performance. However, such
Intelligence and Computing, Tianjin University, Tianjin 300350, China (e-mail: a mechanism takes into account the issue of device energy
[email protected]). consumption, and there is relatively little research on ensuring
Xiuzhen Cheng is with the School of Computer Science and Tech-
nology, Shandong University (SDU), Jinan 250100, China (e-mail: message performance. Although previous joint optimization
[email protected]). methods [17], [18] can minimize system energy consumption
This article has supplementary downloadable material available at while meeting different latency requirements of users. However,
https://doi.org/10.1109/TSC.2025.3539199, provided by the authors.
Digital Object Identifier 10.1109/TSC.2025.3539199 due to the high time complexity, their work cannot be applied

1939-1374 © 2025 IEEE. All rights reserved, including rights for text and data mining, and training of artificial intelligence and similar technologies.
Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Indian Institute of Information Technology Design & Manufacturing. Downloaded on May 16,2025 at 04:55:41 UTC from IEEE Xplore. Restrictions apply.
 840 IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. 18, NO. 2, MARCH/APRIL 2025

to real scenes. Therefore, in actual scenarios, it is a key issue to offloading scheme based on MEC is designed in Section V. A
maximize the energy consumption of the system, while ensuring data center network RCube with heterogeneous edge servers is
the reliability of messages and reducing transmission delay [9]. used for real deployment in Section VI. Performance evaluation
The last but not least challenge comes from effectively de- and analysis are presented in Section VII. The related work is
ploying and managing the server resources required by differ- presented in Section VIII. The final Section concludes the full
ent applications in the edge datacenter [6]. Under the limited paper.
resource capacity of the data center, a reasonable resource de-
ployment strategy is formulated to improve resource utilization, II. MOTIVATION AND OBSERVATION
reduce resource costs, and achieve load balance. Most of the
existing researches focus on workflow task scheduling in cloud A. A Motivation Example
computing environments in special forms of independent or meta This section gives an example to illustrate the pros and cons
tasks [19]. However, previous work cannot be directly applied of both awareness and offloading. A workflow has a certain
here because they overlooked the data association and priority number of tasks, each task processing a specific file. Each file
constraint relationships between tasks, which cannot reflect the has different data volume and complexity, and there are certain
actual characteristics of application tasks. Therefore, it is still dependencies between these files, so the data center should
an unsolved task to make full use of Microservices workflow to analyze this relationship reasonably. We try to minimize the
improve application flexibility. movement and transmission of data during process execution to
In this paper, by joint considering the above three challenges, improve the performance of the data center. A simple example
we propose CACO, a novel framework for jointing optimization of offload message delivery is illustrated in Fig. 1.
of heuristic service deployment and task offloading for data Fig. 1(a) is an example of a simple workflow. Among
centers with edge heterogeneous servers. CACO first establishes them, there are four input files f1 , f2 , f4 , f5 , two out-
a custom adaptive service quality calculation model, which can put files f3 , f6 , and four tasks t1 , t2 , t3 , t4 correspond
overcome the problems that the description of service quality to datasets F S = {f1 , f2 , f3 , f4 , f5 , f6 }, and task sets F =
is not comprehensive enough and the accuracy is not enough. {t1 , t2 , t3 , t4 }. The specific data content is: f1 = 200, t1 , f2 =
CACO next defines the service discovery model of the affinity 400, t1, f3 = 100, t1 , t2 , f4 = 500, t3 , t2 , f5 = 300, t4 ,
matrix, which used to solve the problems of inconvenience f6 = 800, t4 , t2 .
and low accuracy in the dependency graph generation scheme. The relationship between the dataset and tasks during work-
CACO finally presents a content-aware classification offloading flow operation is not one to many or many to one, but rather
algorithm based on MEC. many to many. From Fig. 1(a), f4 is used simultaneously by
We highlight main contributions as follows. tasks t2 and t3 , while task t2 uses both f4 and f3 files. Due
r We propose a QoS calculation model with custom pa-
to the correlation between data in the workflow, closely related
rameters for the calculation of service quality, which can datasets should be placed in the same cluster as much as possible.
evaluate the running status of the service. An adaptive In Fig. 1(b), if the workflow is assigned to three clusters and
service model is designed to achieve the goal of service deployed and scheduled according to the data, f3 is required for
error tolerance and throughput improvement in the service t2 ∈ C2 , while f3 ∈ / C2 and f3 ∈ C1 are generated by t1 . That
running process. is, f3 = {100, t1 , C1 , link = out}, so we need to transfer f3 to
r We investigate a content-aware classification offload al-
C2 . For the same reason, it is necessary to transfer f6 to C3 and
gorithm based on edge computing to balance latency and f4 to C2 . Therefore, a total of three data moves are required,
energy loss. CACO designed an association matrix for with a data volume of 100 + 500 + 800 = 1400.
service discovery algorithms, which can reduce the cost If data deployed by Fig. 1(c), t1 and t2 need to handle f1 ,
and latency of related service calls. In addition, CACO f2 , f3 , f4 , while {t1 , t2 } ∈ C1 , {f1 , f2 , f3 } ∈ C1 , so there is no
deploys an effective priority task scheduling algorithm on need to move data. But f4 ∈ C2 , and t2 also needs to handle f4 ,
edge devices, which has good reduction in latency and so f4 should be transmitted to C1 . Similarly, t4 needs to handle
energy consumption. f5 and f6 , while t4 ∈ C3 and f5 ∈ C3 . But f6 ∈ / C3 , so we need
r Experimental and simulation results show that CACO sig-
to transfer f6 from C1 to C3 , t3 needs to handle f4 , but t3 ∈ C2 ,
nificantly reduces the total traffic of redundant messages f4 ∈ C2 . So there is no need to move data, so it only needs to be
by 46.2% –76.5% . We observed a significant reduction moved twice. The amount of data moved is: 500+800=1300.
in mean task completion time compared to representative
scheduling algorithms. Testbed results also show that our
algorithm can make a trade-off between message process- B. Our Intuition
ing delay and energy loss. CACO also demonstrates its We observe that the two solutions of data deployment and
scalability, reducing control overhead by 34.5% –81.6% scheduling have advantages and disadvantages. We can clearly
compared to state-of-the-art solutions. see that different deployment methods and task scheduling
The rest paper is structured as follows. Motivation and ob- strategies have different impacts on the number of data move-
servation are presented in Section II. The problem formulation ment generated by the workflow. For each task to be executed,
and system model are proposed in Section III. Section IV gives the required data must be located in the same cluster, otherwise
the algorithm design for CACO. Content aware classification the data will need to be transferred from other clusters. If some

Authorized licensed use limited to: Indian Institute of Information Technology Design & Manufacturing. Downloaded on May 16,2025 at 04:55:41 UTC from IEEE Xplore. Restrictions apply.
 XIAO et al.: JOINT SERVICE DEPLOYMENT AND TASK OFFLOADING FOR DATACENTERS WITH EDGE HETEROGENEOUS SERVERS 841

Fig. 1. (a) An example of workflow. (b) Data deployment and scheduling scenario 1. (c) Data deployment and scheduling scenario 2.

datasets are always used by the same task set, there is reason
to believe that deploying these datasets to the same cluster and
then scheduling corresponding tasks to this cluster will reduce
the amount and frequency of data movement, claiming that
these datasets have dependencies. If two datasets are used by the
same task at the same time, it indicates that these two datasets
are related. If more tasks use data sets, the larger the data, and
the greater the correlation between data sets.
A question that immediately arise after the above discussion
is that can we improve the performance of the data center
by minimizing the movement and transmission of data during
process execution? Obviously, we should use data deployment
and task scheduling based on maximum correlation. However,
replicating across multiple sites can lead to data transmission
Fig. 2. Adaptive service discovery model in MEC.
issues [12]. Therefore, it is necessary to effectively reduce the
average completion time of processing request tasks. In addition,
each file has a different amount of data and complexity, and there
nodes contained in the subgraph G(S), which is the total number
are certain dependencies between these files. In other words,
of resources of the edge cloud node combination corresponding
analyzing data correlation and reducing data transmission is
to the subgraph G(S). Let T (G) = {S1 , S2 , . . . , Sm } be the
more intuitive and efficient than existing solutions.
set of subgraphs which composed of all the subgraphs in the
Five servers are depicted as a graph with five nodes, where
graph G.
nodes {A, B, C, D, E} represent independently computable
subtasks. Each edge represents the dependency relationship
B. Network Model
between subtasks, and the weight represents the amount of data
transmitted between subtasks. 1) Adaptive Service Model: Suppose that there are M mobile
device, each terminal generates i security messages, where
III. SYSTEM MODEL M = {M1 , M2 , . . ., Mi }, i = {1, 2, . . ., I}. The security mes-
sage generated by the service request can be modeled by a
In this section, we introduce the adaptive service model,
set of parameters, Mi = {bi , Ci , Ti , Pi }, where bi , Ci , Ti and
correlation matrix, and design objectives.
Pi are the size of the message data, the CPU cycle required
for the message, the deadline, and the priority of the message,
A. Terminologies and Notions respectively. See Fig. 2.
Let graph G = (V, E, w, d) be the network topology of edge The running status can be evaluated through the following key
cloud nodes, which has four attribute elements. V denotes the QoS indicators:
set of nodes of graph G, V = {v1 , v2 , . . ., vn }, each element in Response Time (RT): The average response time of a service
V denotes an edge cloud node in the edge cloud system, then to a request is defined as follows:
the dimension of V is equal to the number of edge cloud nodes.
1 
N
E denotes the set of edges in the graph G, let E = {eij , 0 < RT = (tend,i − tstart,i ) , (1)
i < n, 0 < j < n}, where i and j be the i and j elements in V , N i=1
respectively, and eij is the edge between the vertex vi and vj .
where tend,i is the end time of i-th request, tstart,i is the start
Let w be the weight of the nodes in the graph G, which indicates
time of t-th request, N is the total number of request.
the number of current idle resources of the edge cloud node
Success Rate (SR): The rate at which a service successfully
corresponding to the nodes, such as available CPUs, etc. Let d
processes requests.
be the weight of the edge in the graph G. We define the subset
S ∈ V , and let G(S)  = (S, E ∩ S × S) be a subgraph of the S
SR = , (2)
graph G. Let W (S) = vi ∈S wi be the sum of the weights of the T

Authorized licensed use limited to: Indian Institute of Information Technology Design & Manufacturing. Downloaded on May 16,2025 at 04:55:41 UTC from IEEE Xplore. Restrictions apply.
 842 IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. 18, NO. 2, MARCH/APRIL 2025

where S is the number of successfully processed requests, and the message, the CPU cycles required for the message, and
T is the total number of requests. the deadline requirement. The deadline requirement is more
Resource Utilization (RU): The proportion of resources used important than the size of the message data and the CPU cycle
by a service to the available resources. required for the message. The CPU cycles required for the
Rused message are more important than the size of the message data.
RU = , (3) Therefore, the deadline requirement has the highest weight for
Rtotal
the division of priorities in the analytic hierarchy model. We first
where Rused is the amount of resources used by the service, and compare the factors at the same level in pairs, and then construct
Rtotal is the total available resources. the analytic hierarchy matrix A = (αij )n×n , where
Error Rate (ER): The rate at which errors occur when a service  1
processes requests. = n, i = j,
αij = αji (10)
E 1, i = j.
ER =, (4)
T where αji is the element in comparison matrix A, representing
where E is the number of requests with errors, and T is the total the relative importance between factor i and factor j.
number of requests. Then, we use calculate the comparison matrix weight vector
Availability (AV): The proportion of time a service can be by summation. The vector calculation equation Urk correspond-
accessed within a certain period of time. ing to the weight is:
Tup n
AV = , (5) j=1 αrj
Ttotal Ur = n n
k
, (11)
i=1 j=1 αij
where Tup is the time when the service is running normally,
Ttotal is the total time. where k is the number of influencing factors considered in the
The QoS calculation model for custom parameters can be decision-making process number, r is the message generated by
expressed as: the service request.
Then the vectors corresponding to the weights of all messages
QoS = w1 · Rt + w2 · (1 − SR) form a matrix:
+ w3 · RU + w4 · ER + w5 · (1 − AV ), (6) ⎛ 1 ⎞
u1 u21 u31
⎜ u1 u2 u2 ⎟
where w1 , w2 , w3 , w4 , and w5 , are custom weight parameters ⎜ 3⎟
Φ=⎜ 2 2
⎟ (12)
used to adjust the importance of different indicators. ⎝. . . . . . . . .⎠
Service fault tolerance can be described by the following
u1i u2i u3i
formula:
Fault Recovery Rate (FRR): The rate at which a service where uki is an element in the weight vector Urk , representing
recovers from a fault is defined as follows: the weight of the i-th factor relative to other factors.
R Then we get the eigenvalue corresponding to its weight ac-
F RR = , (7)
F cording to the analytic hierarchy process matrix, denoted by Θ,
where R is the number of successfully restored service instances, Θ = [γ1 , γ2 , γ3 ]T , where
and F is the number of failed service instances. k
1
k
Fault Isolation Rate (FIR): The rate at which a fault service j=1 αrj
γk =  , (13)
is successfully isolated. k i=1 kj=1 αij uj
I where γk is an element in the eigenvalue vector Θ, representing
F RR =
, (8)
F the eigenvalue of the k-th influencing factor.
where I is the number of successfully isolated fault service Finally, we can get the priority vector of each message. Let 
instances, and F is the number of service instances that have be the priority vector and each element in the vector  represents
experienced faults. the priority value of the message. It can be calculated as:
Throughput (TP): The total number of requests processed by ⎛ 1 ⎞
the system per unit of time. u1 u21 u31
⎜ u1 u2 u2 ⎟
⎜ 3⎟
TP =
T
, (9) =Φ×Θ= ⎜ 2 2
⎟
Ttotal ⎝. . . . . . . . . ⎠
where T is the total number of requests processed within time u1i u2i u3i
Ttotal . ⎛ 3 k
⎞
k=1 u1 γk
Suppose that between the time 0 and t, each node gener- ⎜ 3 u k γ ⎟
ates an independent security message M = {M1 , M2 , . . ., Mi }, ⎜ k⎟
×[γ1 , γ2 , γ3 ] = ⎜ k=1 2 ⎟
T
(14)
i ∈ I, i = {1, 2, . . ., I}. When determining the priority of a ⎝ ... ⎠
3 k
message, we mainly consider three factors: the data size of k=1 uj γk

Authorized licensed use limited to: Indian Institute of Information Technology Design & Manufacturing. Downloaded on May 16,2025 at 04:55:41 UTC from IEEE Xplore. Restrictions apply.
 XIAO et al.: JOINT SERVICE DEPLOYMENT AND TASK OFFLOADING FOR DATACENTERS WITH EDGE HETEROGENEOUS SERVERS 843

2) Service Association Matrix: In this section, we build a The maximum amount of dependencies only considers neigh-
service affinity matrix to reduce the number of traversing the boring neighbors, thus causing large data to be clustered together
registry to find related services. This reduces the latency and and small data to be clustered together. Since the correlation
cost of the entire service discovery and invocation process. matrix is symmetric, the simplified correlation is defined as
We first define some basic concepts to better analyze resource follows:
deployment and task scheduling.

n 
n
Definition 1: Let the set C = i=1,2,...,n {ci } be a collection AM = af f (Ai , Aj ) [af f (Ai , Aj−1 )
of clusters, and C denote the set of servers. where ci = capi  i=1 j=1
is the cluster numbered i, and capi is the processing capability
of ci . +af f (Ai , Aj+1 )] , (17)
Due to the fact that a workflow consists of many tasks, each
task requires a certain amount of processing time. Use a specific The correlation matrix is established based on the maximum
numerical value to represent the size of the cluster’s processing amount of correlation between the data. The quantitative cal-
power, which depends on the system parameters during deploy- culation of the maximum correlation amount is carried out
ment. Due to the fact that users can apply for storage space in the by (17). Next, we construct a joint optimization model for
data center on demand, we did not consider the issue of storage latency and energy consumption based on the correlation matrix.
size in the data center, and believe that the data center can meet Let S = {s1 , s2 , . . ., sn } be a set of service nodes, each pro-
the storage needs of users. viding different services. U = {u1 , u2 , . . ., un } is a set of user
Definition 2: Let F S = {f1 , f2 , . . ., fn } be the dataset is the nodes, each requiring different services. There is an association
set of files that the tasks of the workflow need to process. Some matrix AA between service nodes and user nodes, where aij is
of these data files are input data and some are output data. the degree of association between user node ui and service node
Definition 3: Let Fi = {t1 , t2 , . . .ti } be the set of tasks in the sj . The correlation degree can be calculated based on factors
workflow, ti =< runtime, fi >, runtime is the running time of such as service quality, distance, bandwidth, etc. The latency
each task, and fi is the file to be processed by task i. and energy consumption of each service node sj are represented
Definition 4: Let fi = sizei , Fi , dci , linki  be the dataset by tj and ej , respectively.
of the file number i in the workflow, where sizei represents the The objective function is to minimize the weighted sum of
size of the file, Fi = {t1 , t2 , . . .ti } is the set of tasks that process total latency and total energy consumption when all user nodes
the file fi , dci is the cluster to which the file is assigned, and access the service:
linki = {in, out} represents whether the file is input or output.  
minF (T ) = (wt ∗ aij ∗ tj + we ∗ aij ∗ ej ) ∗ xij ,
Definition 5: Let trf = {f1 , f2 , . . ., fn } be the set of files
ui ∈U nj ∈S
needed to complete task t.
(18)
Definition 6: Let tgf = {f1 , f2 , . . ., fn } be the set of inter-
mediate files or output files generated by executing task t.
where F (T ) is the weighted sum of total latency and total energy
Definition 7: Let tim be the set of files moved to perform
consumption, wt and we respectively represent the weight coef-
task i:
⎧ ficients of latency and energy consumption, and wt + we = 1.
⎨t = t − t ( 
n
tkgf ) − trf f, And xij is the indicator variable for user node ui to access service
im irf if r
k=1,k=i node nj . If accessed, xij = 1; otherwise, xij = 0.
⎩
ti ∈ Cl , f ∈ Cl , Cl is the l − th cluster.
Taking the association matrix (AA) as input, arrange the rows IV. ALGORITHM DESIGN
and columns to form a clustered association matrix (CA). Rows In this section, we introduce CACO with the above design ob-
and columns are based on maximizing global dependencies. jectives. CACO consists of three important components. Firstly,
Global dependencies are defined as follows: CACO calculates the resource queue model and perturbs the

n 
n relevant answer vectors using a newly designed matrix mecha-
AM = af f (Ai , Aj )[af f (Ai , Aj−1 ) nism. Then, CACO quantifies the trade-off between offloading
i=1 j=1 delay and energy consumption, and then minimizes energy
+ af f (Ai , Aj+1 ) + af f (Ai−1 , Aj ) consumption.

+ af f (Ai+1 , Aj )], (15)

A. Resource Queue Mechanism
where n is the number of rows or columns of the matrix, and Next, we consider the second component of CACO, namely
aff(Ai , Aj ) is the similarity of two columns or two rows with resource queue mechanism for true answer vector. Security mes-
respect to other columns or rows. sages is processed by edge nodes with using M/M/s queuing
The constraints are: model, let s be the number of edge servers. Assuming that the
af f (A0 , Aj ) = af f (Ai , A0 ) average arrival rate of message j is a Poisson distribution of
λj , and the computing capability of edge nodes is an exponen-
= af f (An+1 , Aj ) = af f (Ai , An+1 ) = 0, (16) tial distribution of (i = 1, 2, . . ., s), the service intensity of the

Authorized licensed use limited to: Indian Institute of Information Technology Design & Manufacturing. Downloaded on May 16,2025 at 04:55:41 UTC from IEEE Xplore. Restrictions apply.
 844 IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. 18, NO. 2, MARCH/APRIL 2025

M/M/s system is: Algorithm 1: Affinity Matrix Based Service Discovery Al-
γj gorithm.
θji = . (19)
iui Require: Matrix Mk (i, j) with row j and column j, server
To maintain system stability, θji must be less than or equal to number m;
1, and the corresponding equilibrium distribution is: Ensure: Associated service collection Vlist2 .
 1: Initialize();
(iθjk )k
ϕ0 , 0 ≤ k ≤ i − 1, 2: SaveServiceAndAddress 1();
ϕk = θjik! ii (20)
3: Record the number of key value pairs N in list1;
i! ϕ0 , k = i.
4: if m > (N +1)(N +2)
then
 (kθji )k i
N
where ϕ0 = [ i−1k=0 k! + θii i!(1 − θji )]−1 . The waiting 5: for i ∈ Vlist1 do
ji
time of message task j assigned to corresponding edge server i 6: Modify the key-value in list1;
is: 7: end for
8: for i ∈ Mk do
θji
tij = ϕi . (21) 9: for j ∈ Mk do
γj (1 − θji )2 10: SaveServiceAndAddress2();
From 11: end for
 line 1 to 9, it calculates how many tasks each file is used
by Ti Tj . Therefore, it takes O(n2 ) time. From lines 10 to 16, 12: end for
it is calculated that among the two files that are related to each 13: else if m ≤ (N +1)(N
N
+2)
then
other, the largest file data is multiplied by the number of related 14: RoundRobin;
tasks, and assigned to the corresponding value of the relationship 15: Save the associated service and its address in list2;
matrix. This takes O(n) time. Thus the time complexity of the 16: end return list2;
Algorithm 2 is O(n2 ).

B. Algorithm Design for CACO of polling and traversing the registry. The service discovery
algorithm based on the affinity matrix is shown in Algorithm 3.
In the characteristics of the service association matrix de-
In Algorithm 3, a loop is to implement element traversal in
scribed in Section IV-A, only the upper triangle of the element
lines 5 ∼ 7. In addition, each step in a loop takes a constant
value may be marked as 1. So when traversing the service
time. Therefore, it takes O(n) time. In lines 8 ∼ 12, a large loop
association matrix, only the upper triangle needs to be traversed.
is used to traverse elements of rows and columns respectively,
The number of traversal services can be denoted as (N +1)(N +2)
.
2 and it runs O(n2 ) times. Moreover, each step in a loop takes
Therefore, the total delay T of service discovery and invocation
a constant time. Thus the time complexity of the algorithm is
can be denoted by:
O(n2 ).
(N + 1)(N + 2) t
T = tr + T R + t k + × , (22)
2 m V. TASK OFFLOADING FOR EDGE HETEROGENEOUS SERVERS
where N + 1 is the row and column size of the service associa- In this section, we investigate task offloading strategies for
tion matrix. The cost C of service discovery and invocation can CACO that are optimal for latency and energy. Considering
be denoted by: the relationship between delay and energy consumption, an
C = CR + C D effective priority task scheduling algorithm is deployed on the
  edge device.
1
n
(N + 1)(N + 2) t
+ Oi × A × tk + × . (23)
n i=1 2 m A. Task Queue Scheduling
a For each edge server i, let i = {1, 2, . . ., I} be the set of
If and only if b=1 ykb > (N +1)(N 2
+2)
, the method of con-
structing the service association matrix is better than the method edge servers that store different message types, let Wi be its
of polling and traversing the registry, where a represents the communication bandwidth, and let Vi denote the maximum
number of associated services of the service Mk . Therefore, a is computing rate. We consider the offloading problem of secure
equivalent to N . Considering that in the general case, assuming messages in period t. From time 0 to t, when the security message
that the order average value of the associated services of service j is unloaded to the edge server i, the communication bandwidth
Mk in the registry is M/2, the inequality can be transformed allocated by the edge server i for the message j is wij , and the
into: allocated computing rate is represented by vij . During cycle
time [1, t], the total required communication bandwidth and
mN (N + 1)(N + 2)
> , (24) computation rate of all tasks offloaded to i should not exceed
2 2 the resource amounts Wi and Vi , respectively.
It can be deduced from the (24) that if and only when Since the message task may not be unloaded immediately, it
m > (N +1)(N
2
+2)
, the method of constructing the service asso- needs to wait for a period of time at the edge node before it is
ciation matrix in the average situation is better than the method transmitted to the corresponding edge server. Therefore, the total

Authorized licensed use limited to: Indian Institute of Information Technology Design & Manufacturing. Downloaded on May 16,2025 at 04:55:41 UTC from IEEE Xplore. Restrictions apply.
 XIAO et al.: JOINT SERVICE DEPLOYMENT AND TASK OFFLOADING FOR DATACENTERS WITH EDGE HETEROGENEOUS SERVERS 845

delay of system offloading includes latency delay, transmission is unloaded to the corresponding edge server i within time [0, t],
delay and computation delay. When message task j is offloaded otherwise xi,j,t = 0, it means that task j will not be offloaded
to edge server i at time period t, the total delay is calculated to edge server i for execution at time t, but will be executed
by (25) as follows: on the local device. The local computation delay dlocal
i,j,t can be

bj cj calculated using the following formulas:

di,j,t = ti,j + + , (25)
ri,j vi,j
dlocal
j,t = cj /clocal , (28)
where ti,j is the waiting time of message task j before the edge
bj cj where cj is the number of CPU cycles required for task j, and
node is scheduled. ri,j is the transmission delay, and vi,j is
the computation delay, which represents the transmission power clocal is the computing power of the local device.
between message task j and edge server i. cj is the offload size The energy consumption of local calculation, plocal
j,t , can be
of the message task. calculated using the following formula:
In the process of computing offloading, the energy consump-
tion of edge nodes generally includes message task offloading plocal
j,t = cj elocal , (29)
transmission consumption and computing processing consump-
tion. We offload the message task to the edge server for pro- where cj is the number of CPU cycles required for task j, and
x
elocal is the power consumption of the local device per CPU
cessing and define a function η(x) = N0 wi,j (2 wi,j − 1), so the
cycle.
total energy consumption can be expressed as:
The total delay Dtotal and total energy consumption Ptotal
bj cj can be calculated using the following formula:
βi,j = 2 η(x) + ei . (26)
ri,j Si,j vi,j  
Dtotal = di,j,t xi,j,t + dlocal
j,t , (30)
where Sij is the channel gain between edge server i and message
i∈I j∈J t∈T j∈J t∈T
task j, and ei is the power consumption of edge server i per CPU  
cycle. Ptotal = pi,j xi,j,t + plocal
j,t . (31)
i∈I j∈J t∈T j∈J t∈T
B. Task Classification Offloading
R5 is the actual energy consumption of the task message, which
In order to explore the relationship between computing of-
must be less than the CPU cycles required for the expected
fload delay and energy consumption, we assign different weight
message.
coefficients to offload delay and energy consumption. The main
δ1 and δ2 are weights of delay and energy loss, and δ1 + δ2 =
problem can be formulated as follows:
1. We can study the relationship between energy loss and delay

(δ1 di,j,t + dlocal by changing the weight coefficients in the objective function.
j,t ) + (δ2 βi,j + pj,t ))xi,j,t ,
local
min
i∈I j∈J t∈T For example, when a larger weight is set for the delay, the main
goal of this problem is to minimize the total delay, and the
s.t. proposed algorithm tends to use larger transmission power to
  allocate tasks. Conversely, when a larger weight is placed on
R1 : wi,j xi,j,m ≤ Wi , ∀i ∈ I, t ∈ T,
the energy loss, the main goal of the problem is to minimize the
1≤m∈t j∈J
  energy loss, then it will tend to use the minimum transmission
R2 : vi,j xi,j,m ≤ Vi , ∀i ∈ I, t ∈ T, power to allocate tasks. Since the problem contains two decision
1≤m∈t j∈J variables di,j,t and pi,j and is the product of the two variables,
 the problem is non-convex. We first transform a non-convex
R3 : xi,j,t = 1, ∀j ∈ J, problem into a convex problem via Lagrangian relaxation [18].
i∈I t∈T Its Lagrangian relaxation function is as follows:
R4 : xi,j,t ∈ {0, 1}, ∀i ∈ I, j ∈ J, t ∈ T, 
Lx,γ = (δ1 di,j,t + dlocal
j,t )+(δ2 βi,j +pj,t ))xi,j,t ,
local
bj cj
R5 : 2 η(x) + v ei ≤ Cj , ∀i ∈ I, j ∈ J. (27) i∈I j∈J t∈T
ri,j Si,j i,j ⎛ ⎛ ⎞⎞
  
Restriction R1 is the sum of the occupied communication + ⎝γit
0 ⎝
wi,j xi,j,m − Wi ⎠⎠
bandwidth of all tasks offloaded to edge server i from time 1 to t, i∈I t∈T 1≤m∈t j∈J
which cannot exceed the accumulated available communication ⎛ ⎞
resources Wi for all periods t. R2 is the sum of the required  
1 ⎝
computing rates of all tasks of edge server i from period 1 + γi,t vi,j xi,j,m ≤ Vi ⎠
to t, which cannot exceed the cumulative available computing 1≤m∈t j∈J

resources Vi for all periods t. R3 states that each task j can be   

 
assigned to exactly one edge server i in only one cycle t. R4 is a + γj2 xi,j,t − 1
binary variable constraint, xi,j,t = 1 means that message task j j∈J i∈I t∈T

Authorized licensed use limited to: Indian Institute of Information Technology Design & Manufacturing. Downloaded on May 16,2025 at 04:55:41 UTC from IEEE Xplore. Restrictions apply.
 846 IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. 18, NO. 2, MARCH/APRIL 2025

Algorithm 2: Task Queue Scheduling Algorithm. Algorithm 3: Message Task Offload.

Require: The data size bj , the CPU cycle Cj , the deadline Require: Task M , edge server set I, allocated
requirement Tj , the priority level Pj ; communication bandwidth wij , allocated calculation rate
Ensure: Task queue list; is vij ;
1: for Mj in edge server do Ensure: Distribution coefficient x and objective function
2: if pj = 3 then value z ∗ ;
3: Place the message Mj in to QH ; 1: for i ∈ I and j ∈ M do
4: A = (ai,j )m×n ; 2: γ 0 , γ 1 , γ 2 and γ 3 , and obtain the transmission power
5: Calculate Urk , Ω = [γ1 , γ2 , γ3 ]T ; by (26);
6:  = Φ × Θ; 3: z ∗ = 0, compute Wi,j and Vi,j ;
7: Arrange QH in order by ; 4: if Wi,j < Wi and Vi,j < Vi then
8: else if pj = 2 then 5: x = 1;
9: Put the message into QM ; 6: else
10: Repeat step 4 to 7; 7: x = 0;
11: Arrange QM in order by ; 8: end if
12: else if pj = 1 then 9: Update the objective function (34) by x;
13: Put the message into QL ; 10: Update the Lagrangian multiplier by the subgradient
14: Repeat step 4 to step 7; projection of ψ(γ), and update the transmission
15: Arrange QL in order by ; power using the KKT condition;
16: end if 11: end for
17: end for
18: return Task queue;

1
   γi,t (t + 1) =
 bj cj ⎡ ⎛ ⎞⎤ +
3
+ γij 2 η(x) + v ei − C j ,  
ri,j Si,j ⎣γi,t
i∈I j∈J
i,j 1
(t) − p1i,t ⎝ vi,j xi,j,m − Vi ⎠⎦ , (36)
(32) 1≤m∈t j∈J
 
where γ = {γ 0 , γ 1 , γ 2 , γ 3 }, γ 0 and γ 1 ∈ ∂ I×T , γ 2 ∈ ∂ J , γ 3 ∈ 
γj2 (t + 1) = γj2 (t) − p2j xi,j,m − 1 , (37)
∂ I×J , γ ≥ 0.
i∈I t∈T
Then the problem is transformed into solving an optimization
problem as follows: γj3 (t + 1)
  +
ψ(γ) = minx L(x, γ). (33) 3 bj cj
= γi,j (t) − p3i,t 2 η(x) + v ei − C j , (38)
ri,j Si,j i,j
It is easy to know that the function ψ(γ) is a convex func-
tion, and ψ(γ) ≤ L(x, p, γ), then the above problem is further where p1i,t , p2i,t , p3i,t and p4i,t are the step sizes corresponding to
formulated as its dual problem as follows: γ0 , γ1 , γ2 , and γ3 . respectively.
The specific flow is shown in Algorithm 4. After obtaining the
maxγ ψ(γ) = maxγ ψ[minx L(x, γ)],
priority corresponding to each message, we first perform task
s.t. scheduling assignment in the message queue according to the
priority of the task. The edge server for message offloading has
γ 0 , γ 1 ∈ ∂ I×T , three priority levels of high, medium and low, namely QH , QM
γ2 ∈ ∂J , and QL . And queue storage is further performed according to
the message priority in each edge server. There are three types of
γ 3 ∈ ∂ I×J , message tasks, namely pj = {1, 2, 3}, where 3 is high priority, 2
γ ≥ 0. (34) is medium priority, and 1 is low priority. Therefore, all messages
with strict latency requirements should be offloaded to high-
The subgradient projection of ψ(γ) is priority edge server queues.
The main problem of offloading adopts subgradient projection
0
γi,t (t + 1) and heuristic greedy algorithm to get a feasible solution after
⎡ ⎛ ⎞⎤ + completing the task queue scheduling. In subgradient projection,
  the value of the Lagrangian multiplier is updated with certain
= ⎣γi,t
0
(t) − p0i,t ⎝ wi,j xi,j,m − Wi ⎠⎦ , (35) steps and the transmission power is also updated according to
1≤m∈t j∈J the KKT condition. Considering the computational complexity,

Authorized licensed use limited to: Indian Institute of Information Technology Design & Manufacturing. Downloaded on May 16,2025 at 04:55:41 UTC from IEEE Xplore. Restrictions apply.
 XIAO et al.: JOINT SERVICE DEPLOYMENT AND TASK OFFLOADING FOR DATACENTERS WITH EDGE HETEROGENEOUS SERVERS 847

TABLE I
THE OPTIMAL COMPROMISE SOLUTIONS AT DIFFERENT p

Fig. 3. RCube(4,1,1) structure: consists of 4 RCube(4,1,0) and 4 switches.

uniformly located within this range, and the number of security
messages generated by the server at the same time does not
we use the greedy algorithm to solve the integer programming exceed 30. Each message has a size of 400 ∼ 10000 Bytes and
problem to obtain the distribution coefficient x and the value of a latency period of 100–2000 ms. The initial priority of the
the objective function. Thus Algorithm 4 can implement the task message is pj = {1, 2, 3}, and the required CPU size for the
offloading strategy in polynomial time. message is 1 × 106 to 6 × 106 cycles/s, security messages arrive
at the signal unit in a evenly distributed manner. The distance
VI. TASK DEPLOYMENT IN RCUBE NETWORK between the signal unit and the edge server is 100 ∼ 1000 meters,
and the transmission path loss is 3 dB. The communication
In this section, we implement the application of CACO in the bandwidth of the edge server is 1 ∼ 3 MHz, and the calculation
RCube network. Rcube is a highly scalable and energy-efficient rate is 1 ∼ 15 × 108 cycle/s. We study the relationship between
DCN, which improves the availability of the network by deploy- energy loss and delay by changing the weight coefficients in the
ing an active/passive redundancy scheme. objective function. δ1 and δ2 are weights of delay and energy
loss, and δ1 + δ2 = 1. Including the following: 1) δ1 = δ2 =
A. RCube Network 0.5; 2) δ1 = 0.3 and δ2 = 0.7; 3) The variation between delay
In RCube(n, m, k), RCube(n, m, 0) is its underlying struc- and energy loss under different settings of δ1 .
ture, which consists of switches with n + m servers connecting
(n + m)-ports. Meanwhile, the servers in RCube are divided
B. Simulation Evaluation
into core servers and edge servers. In an RCube(n, m, 0), there
are n core servers and m edge servers. First, we evaluate the performance of the data perturbation
RCube(n, m, k) is composed of n number of RCube mechanism for CACO.
(n, m, k − 1) and nk number of (n + m)-port switches. Fig. 3 Let the p-value be the target significance level, ηex be the
shows the RCube(4,1,1). BCube is a special case of RCube, efficiency, pi,j be the energy cost. Let 1 − α be the uncertainty
when m = 0, RCube(n, 0, k) can be regarded as BCube(n, k). confidence level. Due to the impact of different levels of target
For the same size network, RCube consumes less energy than significance on system scheduling results, p-values of 0.05, 0.15,
BCube without any energy-aware routing algorithm. In addition, and 0.25 were set for optimization and solution. The frontier of
RCube is configured with active/passive redundancy mode, so the obtained objective function is shown in Fig. 8, and Table I
that the performance of the server in the standby state is not presents the optimal compromise solutions at different levels of
affected, which further enhances the availability of the network objective significance. From Fig. 8 and Table I, it can be seen that
significantly. the operating cost and efficiency of CACO are interdependent.
Please refer to Appendix A for the task deployment of RCube This is because in order to achieve higher efficiency, the demand
network with load balancing. for resources will inevitably increase, leading to an increase in
the operating costs of CACO. The smaller the p-value setting, the
VII. SIMULATION AND EXPERIMENTS lower the leading edge curve of the objective function will be.
The higher the probability that the operating cost and efficiency
In this section, we conduct large-scale simulation experiments
of CACO are better than expected. But the smaller the p-value,
in CloudSim and deploy actual algorithms on a real comput-
the lower the confidence level and the weaker the robustness.
ing cluster. Our laboratory (Nanjing Branch of the National
Therefore, the system should set an appropriate p-value based
High Performance Computing Center, NHPCC), which has 300
on its own risk tolerance.
types of servers, including high-performance computing servers,
1) Evaluation of Latency and Security Message: We com-
domestic Loongson servers, storage servers, GPU servers, AI
pare CACO with the Round-Robin [32] and the functional
artificial intelligence servers, etc.
decomposition heuristic service discovery (FDHSD) [35] in the
total delay T and cost C.
A. Performance Metrics and Benchmarks Fig. 4(a) shows the variation of system latency and
We evaluate the performance of the proposed algorithm on service quantity under different weight conditions δ1 (δ1 =
the CloudSim platform [5]. In the MEC scenario, the communi- 0.3, 0.5, 0.7). As the weight δ1 increases, the system latency
cation range of each signal unit is 1000 m. All mobile nodes are is expected to decrease under the same service volume. When

Authorized licensed use limited to: Indian Institute of Information Technology Design & Manufacturing. Downloaded on May 16,2025 at 04:55:41 UTC from IEEE Xplore. Restrictions apply.
 848 IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. 18, NO. 2, MARCH/APRIL 2025

Fig. 4. (a) Performance comparison of latency. (b) Comparison of energy cost. (c) Comparison of service cost under different strategies. (d) Comparison of
latency under different strategies.

Fig. 5. (a) The relationship between latency and cost. (b) The impact of changes in z on uninstallation decisions. (c) Comparison of system availability.
(d) Comparison of delivery success rate.

the weight value δ1 increases, reducing system latency be- Fig. 5(a) shows that the larger the value of δ1 , the higher
comes the main optimization objective. CACO adapts to the the delay metric requirement of the proposed algorithm. The
goal of reducing latency by optimizing the selection of tasks trend of delay variation under different δ1 values, with the target
for computation on local devices or offloading to the edge and value being the weighted sum of average delay and energy loss.
cloud. Fig. 4(b) shows the changes in total energy consump- As δ1 increases from 0.001 to 0.999, the delay metric gradu-
tion of mobile terminals in the system under different weights ally dominates the value of the objective function. Therefore,
δ1 (δ1 = 0.3, 0.5, 0.7). For CACO, the total energy consumption energy loss will increase due to the reduction of weight δ2 in
of mobile terminals in the system increases with the weight. the objective function. Let z represent the ratio between task
When δ1 increases, reducing system latency becomes the main computation requirements and communication requirements.
objective of optimization. As the number of services and tasks Fig. 5(b) shows the task offloading situation under different
increases, the computing tasks of edge nodes and cloud nodes ratios of computing and communication requirements z. When
become heavier. As the computation time increases, the ex- z < 4, no uninstallation occurs. As z increases, the amount
ecution time of mobile terminals becomes relatively shorter. of uninstallation gradually increases. After z > 20, tasks are
Therefore, the proportion of mobile terminals in task allocation basically offloaded to the edge and cloud. Due to the small ratio,
will correspondingly increase. mobile terminals can meet resource requirements. When the
In Fig. 4(c), δ1 = 0.5. Compared with the round robin and ratio gradually increases and the mobile terminal cannot meet the
function decomposition heuristic methods, CACO has lower resource requirements, it will be offloaded to remote execution.
latency. As the workload continues to increase, the compu- Fig. 5(c) shows the process of increasing the concurrency
tational difficulty of tasks also increases.CACO has a slower from 50,000 to 120,000. the microservice system imple-
growth rate compared to other strategies. Reasonable allocation ments a service correlation-based hierarchical degradation al-
of resources for edge nodes can effectively reduce the task gorithm (SRBHD), an automated timeout degradation algo-
execution time of the system when implementing computation rithm (ATDA), and an automated current limiting degradation
offloading. Setting δ1 = 0.3, Fig. 4(d) shows the service cost algorithm (ACLDA). When the concurrency is about 70,000,
of the system. CACO considers message processing latency and the system availability drops rapidly after the system executes
energy consumption, and has developed a reasonable strategy to the automatic current limiting degradation algorithm. However,
determine the priority of secure messages. Deploy task priority after the system executes the CACO, the system availability
scheduling algorithms in edge servers, with each edge server decline rate is relatively slow, and the system availability is as
handling security messages of different priorities. CACO uses high as 90% or more in most of the time.
slightly higher power than the minimum transmission power to Fig. 5(d) shows the delivery success rates of the four algo-
transmit messages, thereby reducing its computation delay. rithms. The successful delivery rate refers to the ratio of the

Authorized licensed use limited to: Indian Institute of Information Technology Design & Manufacturing. Downloaded on May 16,2025 at 04:55:41 UTC from IEEE Xplore. Restrictions apply.
 XIAO et al.: JOINT SERVICE DEPLOYMENT AND TASK OFFLOADING FOR DATACENTERS WITH EDGE HETEROGENEOUS SERVERS 849

Fig. 6. (a) Comparison of algorithm service cost. (b) Comparison of service cost and the number of security messages. (c) Task completion time for different
numbers of mobile devices. (d) Task completion time under different number of tasks.

number of successfully received data packets to the total amount From Fig. 6(c) and (d), regardless of the size of the task graph,
of transmitted data within the specified time. Since CACO con- CACO saves significantly more computational costs than ATDA
siders the relationship between delay and energy consumption, and ACLDA. Similarly, regardless of the density of the task
an effective priority task scheduling algorithm is implemented graph, CACO can provide the optimal offloading strategy by
at the edge device end. comparison.
2) Evaluation of Completion Time: As shown in Fig. 7(a),
C. System Implementation with the continuous increase of the input data, the completion
time of the method between 500 MB and 1000 MB is signifi-
1) Implementation on the Platform: In general, we use 10
cantly different. For CACO, the system calculation cost, com-
servers running Ubuntu 18.04 and Linux kernel 5.4 to build test
munication and waiting time cost are fully considered. As shown
beds. All servers are equipped with a 22 core Intel Xeon 6152
in Fig. 7(b), as the number of application services increases, the
processor, 128 GB of memory, and Intel X710 10GbE NIC.
average completion time of the scheduling method that is only
We refer to the deployment of the graph processing platform
executed locally is significantly higher than other methods.
GraphLite, which includes two computing clusters. Cluster 1 is
Fig. 7(c) shows the total time cost of ATDA and CACO will
used to simulate edge device, and cluster 2 is used to simulate
decrease as the number of nodes increases. The CACO makes
cloud computing centers. The bandwidth between edge device
good use of the increased computing node resources and is
is 100 Mbps, and the bandwidth between cloud center servers is
significantly better than other methods in terms of processing
30 Gbps. Set the uplink bandwidth to 100 Mbps and the downlink
efficiency. As shown in Fig. 7(d), gradually increasing the band-
bandwidth to 200 Mbps.
width of the entire system has the greatest impact on the method
We select the PageRank algorithm used by Google as the
of communicating only to the cloud center. When the bandwidth
offloading object for the computing task. We select three
is increased to 100 Mbit/s, the communication time cost brought
datasets of different sizes from [20], [21], [22], each containing
by communication transmission is already very small.
15763325729 and 685230 pieces of data. For simplicity, we will
3) Evaluation of Recall Ratio: Fig. 9(a) shows the recall
refer to it as dataset 1, dataset 2, and dataset 3. Meanwhile,
rate of discovery results as the number of services increases.
in order to better compare the performance of the algorithms
CACO comprehensively considers service context factors and
proposed in this study, we also selected graph datasets of differ-
user context factors, and is superior to the other two methods
ent sizes from KONECT, which have different sizes and edge
in terms of recall and precision of service results. Fig. 9(b)
densities.
shows the change in discovery accuracy as the number of
services increases. Due to the existence of multiple service
D. Testbed Results types, some services express different meanings in different
1) Evaluation of Communication Costs: Fig. 6(a) shows the environments, which leads to the low accuracy of microservice
comparison of the load rate of the microservice system executing discovery algorithms that do not consider contextual semantics.
CACO, ATDA [37], and ACLDA. In the process of increasing the Fig. 9(c) shows that as the number of microservices increases,
concurrent amount from 50,000 to 120,000, the smaller the value the time overhead of the four microservice discovery methods
of the y-axis load, the better the system load condition, and the has increased. When the number of test sample microservices
larger the value, the worse the system load condition. Fig. 6(b) is less than 600, the discovery time of SRBHD algorithm is
shows the process of increasing concurrency from 50,000 to longer than that of ACLDA. Fig. 9(d) shows that the priority
120,000, the larger the y-axis value, the more total completion of each message is different, and its expectations for message
time (TCT) that the microservice system can read and write. processing delay are different, as the message priority level
When the concurrency of the CACO is about 70,000 and the increases.
concurrency is about 90,000, the y-axis corresponding system’s 4) Evaluation of Load Balance: As shown in Fig. 10(a), as
ability to read and write becomes larger. the number of tasks increases, the node load increases, and the
We apply the algorithm to three task graphs of different node load balance degree (NLBD) of each framework increases
sizes, with a number of task nodes ranging from 50 to 200. to varying degrees. Comparing and analyzing the mean NLBD

Authorized licensed use limited to: Indian Institute of Information Technology Design & Manufacturing. Downloaded on May 16,2025 at 04:55:41 UTC from IEEE Xplore. Restrictions apply.
 850 IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. 18, NO. 2, MARCH/APRIL 2025

Fig. 7. (a) Average completion time varies with task size. (b) The average completion time varies with the type of task. (c) The average time cost of the algorithm
changes with the number of mobile computing nodes at the edge. (d) Comparison of the impact of bandwidth on the three algorithms.

VIII. RELATED WORK

Adaptive service model: Most existing work have focus on
the adaptability of service composition in dynamic distributed
networks. Chen et al. [23] proposed a service association rule to
improve the quality of service discovery by shorten the delay of
service discovery and increase the success rate. Shi et al. [25]
proposed an adaptive service composition model based on ser-
vice quality, but it is too complex to handle a large number of
Fig. 8. Front edge of robust optimization scheduling objective function based services and QoS attributes. CACO can more effectively handle
on CACO under different p-values. service dynamics and environmental uncertainty. Jia et al. [40]
studied a QoS aware offload framework in serverless MEC,
but they mainly focused on the balance of energy consumption
under different tasks, Round Robin is the largest, followed by and delay. CACO can effectively reduce latency and energy
ATDA and ACLDA, and CACO is the smallest. ATDA randomly consumption by constructing service association matrices and
offloads tasks to the edge or cloud center, with significant priority task scheduling algorithms. Caporuscio et al. [33] pro-
differences in node load rates and a higher NLBD in the range of posed the use of machine learning techniques in the process of
[0.2255 0.4698]. As shown in Fig. 10(b), as the number of edge service discovery. Vaidhyanathan et al. [34] combined machine
devices increases, their load decreases, and the NLBD of the learning and adaptive techniques to perform service discovery
algorithm decreases to varying degrees. The NLBD of ATDA and selection by balancing different QoS attributes.
decreased from 0.5505 to 0.3072, but there was no optimization Task deployment: Most of the existing MEC models perform
task offloading, and its average NLBD was higher, reaching coarse-grained task migration for the tasks to be processed [31].
0.4334. Xia et al. [30] presented a framework based on rounding and
As shown in Fig. 10(c) and (d), as the number of terminals dynamic programming to ensure service availability through the
increases, the average response rate of each algorithm gradually use of VNF extended fault elimination (SAFE). Li et al. [28] pro-
decreases, while the overload probability gradually increases. posed a distributed scheduler TapFinger, which minimizes the
This is because MECs have limited bandwidth and computing total completion time of ML tasks in multi cluster edge networks.
resources. Overall, when the number of terminals reaches its Zhao et al. [38] used the deep learning method to optimize the
maximum, ACLDA and Round Robin perform poorly. Com- reliability and delay in edge computing, but lacked consideration
pared to a terminal count of 30, ATDA, ACLDA, and Round of service dynamics and priority constraints. CACO can adapt
Robin decreased their average response rates by approximately more flexibly to application requirements and prioritize urgent
3.9% , 87.2% , and 36.5% , respectively. and high priority tasks.
5) Evaluation of Load Balance: As shown in Fig. 11(a), However, the above work ignored the correlation between data
when the task size remains unchanged and computing resources and did not consider fine-grained resource allocation of tasks or
decrease, the average offloading utility related to latency and joint optimization of task placement and resource allocation.
energy consumption will correspondingly decrease significantly. Firstly, CACO’s adaptive service discovery model can better
As shown in Fig. 11(b) and (c), the delay and energy con- adapt to the dynamic changes in the edge environment. Secondly,
sumption required to complete the task will naturally be greatly while ensuring service reliability, CACO’s priority offloading
reduced. As shown in Fig. 11(d), due to the decrease in the strategy can improve the efficiency of task processing.
required delay to complete the task, the average response rate Task offloading: More related work have concentrated on
gradually increases and reaches its maximum value. Overall, the task offloading of MEC [29]. Gu et al. [16] explored the
CACO can maintain low latency and energy consumption, while (edge) neighbor connectivity of two types of networks: bicubic
ensuring relatively high average offloading utility and average networks (including bicubes) and generalized commutative hy-
response rate. percubes (including commutative hypercubes). They conducted

Authorized licensed use limited to: Indian Institute of Information Technology Design & Manufacturing. Downloaded on May 16,2025 at 04:55:41 UTC from IEEE Xplore. Restrictions apply.
 XIAO et al.: JOINT SERVICE DEPLOYMENT AND TASK OFFLOADING FOR DATACENTERS WITH EDGE HETEROGENEOUS SERVERS 851

Fig. 9. (a) Comparison of recall rate of discovery results. (b) Comparison of accuracy of findings. (c) Comparison of microservice discovery time. (d) Relationship
between average delay and message priority.

Fig. 10. (a) NLBD for different tasks. (b) NLBD for different edge devices. (c) Comparison of overload probability. (d) Comparison of average response rate.

Fig. 11. (a) Comparison of Average offloading utility. (b) Comparison of average latency. (c) Comparison of average energy cost. (d) Comparison of average
response rate.

simulation experiments to evaluate the disruption effect of dif- IX. CONCLUSION

ferent disruption vertex selections in these networks at different In this paper, we propose a task offloading framework CACO
scales. Tao et al. [36] proposed an energy optimized offloading
for service deployment based on content aware classification
scheme, which is superior to local calculation and complete offloading. In CACO, security messages are prioritized using
offloading methods in terms of energy consumption and delay analytic hierarchy process, and an optimal task offloading model
performance. Ma et al. [39] presented a resource constrained
for latency and energy loss is established. In addition, based on
random correlation rounding algorithm. They round the score the maximum correlation model, the data in the workflow is
variables appropriately to obtain feasible but near optimal so- extracted to form an association matrix. We have evaluated the
lutions based on the physical resource capacity of each edge
performance of CACO through experiments based on real data.
server. Chen et al. [1] attempted to solve the joint optimization Evaluation and analysis indicate that CACO has achieved a sat-
problem of edge application layout (EAP-CR) coverage and
isfactory compromise between message processing latency and
reliability. Ma et al. [24] proposed a credibility decision strategy
energy loss within an acceptable operating time, while ensuring
in response to the high latency of service discovery, which can all expected solution complexity, indicating the usefulness and
reduce repeated trusted calculations under certain conditions.
feasibility of CACO.
Nevertheless, the work does not consider the trade-off be-
tween delay and energy consumption, hence being detrimental
REFERENCES
to resource efficiency. The current optimization research on the
service discovery mechanism has achieved certain results, but [1] F. Chen, J. Zhou, X. Xia, Y. Xiang, X. Tao, and Q. He, “Joint optimization
of coverage and reliability for application placement in mobile edge
it still cannot fully solve the problem of long delay and high computing,” IEEE Trans. Serv. Comput., vol. 16, no, pp. 3946–3957,
cost. Nov./Dec. 2023.

Authorized licensed use limited to: Indian Institute of Information Technology Design & Manufacturing. Downloaded on May 16,2025 at 04:55:41 UTC from IEEE Xplore. Restrictions apply.
 852 IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. 18, NO. 2, MARCH/APRIL 2025

[2] Z. Ning et al., “A cooperative quality-aware service access system for [26] H. Liu, Z. Cao, and X. Zhang, “An efficient algorithm of context-clustered
social internet of vehicles,” IEEE Internet Things J., vol. 5, no. 4, microservice discovery,” in Proc. 2nd Int. Conf. Comput. Sci. Appl. Eng.,
pp. 2506–2517, Aug. 2018. 2018, pp. 1–6.
[3] Z. Yan, P. Cheng, Z. Chen, B. Vucetic, and Y. Li, “Two-dimensional [27] Z. Fan, W. Yang, F. Wu, J. Cao, and W. Shi, “Serving at the edge: An
task offloading for mobile networks: An imitation learning framework,” edge computing service architecture based on ICN,” ACM Trans. Internet
IEEE/ACM Trans. Netw., vol. 29, no. 6, pp. 2494–2507, Dec. 2021. Technol., vol. 22, no. 1, pp. 1–27, 2021.
[4] Y. Liu, Y. Mao, Z. Liu, F. Ye, and Y. Yang, “Joint task offloading and [28] Y. Li, T. Zeng, X. Zhang, J. Duan, and C. Wu, “TapFinger: Task placement
resource allocation in heterogeneous edge environments,” IEEE Trans. and fine-grained resource allocation for edge machine learning,” in Proc.
Mobile Comput., vol. 23, no. 6, pp. 7318–7334, Jun. 2024. IEEE Conf. Comput. Commun. Workshops, 2023, pp. 1–10.
[5] Y. Feng et al., “Evaluating chiplet-based large-scale interconnection [29] J. Ren et al., “An efficient two-layer task offloading scheme for MEC
networks via cycle-accurate packet-parallel simulation,” in Proc. 2024 system with multiple services providers,” in Proc. IEEE Conf. Comput.
USENIX Annu. Tech. Conf., 2024, pp. 731–747. Commun., 2022, pp. 1519–1528.
[6] Q. Peng, Y. Xia, M. C. Zhou, X. Luo, and M. Lin, “Reliability-aware [30] R. Xia et al., “SAFE: Service availability via failure elimination through
and deadline-constrained mobile service composition over opportunistic VNF scaling,” IEEE/ACM Trans. Netw., vol. 31, no. 5, pp. 2042–2057,
networks,” IEEE Trans. Automat. Sci. Eng., vol. 18, no. 3, pp. 1012–1025, Oct. 2023.
Jul. 2020. [31] Y. He, D. Zhai, R. Zhang, J. Du, G. S. Aujla, and H. Cao, “A mobile edge
[7] W. Fan, F. Xiao, J. Fan, Z. Han, L. Sun, and R. Wang, “Fault-tolerant computing framework for task offloading and resource allocation in UAV-
routing with load balancing in LeTQ networks,” IEEE Trans. Dependable assisted VANETs,” in Proc. IEEE Conf. Comput. Commun. Workshops,
Secure Comput., vol. 20, no. 1, pp. 68–82, Jan./Feb. 2023. 2021, pp. 1–6.
[8] W. Fan, F. Xiao, X. Chen, L. Cui, and S. Yu, “Efficient virtual network [32] W.-K. Chung et al., “Dynamic parallel flow algorithms with cen-
embedding of cloud-based data center networks into optical networks,” tralized scheduling for load balancing in cloud data center net-
IEEE Trans. Parallel Distrib. Syst., vol. 32, no. 11, pp. 2793–2808, works,” IEEE Trans. Cloud Comput., vol. 11, no. 1, pp. 1050–1064,
Nov. 2021. Jan.-Mar. 2023.
[9] K. Cao, M. Chen, S. Karnouskos, and S. Hu, “Reliability-aware personal- [33] M. Caporuscio, M. De Toma, H. Muccini, and K. Vaidhyanathan,
ized deployment of approximate computation iot applications in serverless “A machine learning approach to service discovery for microservice
mobile edge computing,” IEEE Trans. Comput.-Aided Des. Integr. Circuits architectures,” in Proc. Softw. Architecture: 15th Eur. Conf., 2021,
Syst., pp. vol. 44, no. 2, pp. 430–443, Feb. 2025. pp. 66–82.
[10] S. Wang, Z. Ding, and C. Jiang, “Elastic scheduling for microservice [34] K. Vaidhyanathan, M. Caporuscio, S. Florio, and H. Muccini, “ML-
applications in clouds,” IEEE Trans. Parallel Distrib. Syst., vol. 32, no. 1, enabled service discovery for microservice architecture: A QoS ap-
pp. 98–115, Jan. 2021. proach,” in Proc. 39th ACM/SIGAPP Symp. Appl. Comput., 2024,
[11] C. Wang et al., “Joint server assignment and resource management for pp. 1193–1200.
edge-based MAR system,” IEEE/ACM Trans. Netw., vol. 28, no. 5, [35] S.-J. Jian and S.-Y. Hsieh, “A niching regression adaptive memetic al-
pp. 2378–2391, Oct. 2020. gorithm for multimodal optimization of the Euclidean traveling salesman
[12] K. Liu, C. Liu, G. Yan, V. C. Lee, and J. Cao, “Accelerating DNN inference problem,” IEEE Trans. Evol. Comput., vol. 27, no. 5, pp. 1413–1426,
with reliability guarantee in vehicular edge computing,” IEEE/ACM Trans. Oct. 2022.
Netw., vol. 31, no. 6, pp. 3238–3253, Dec. 2023. [36] X. Tao, K. Ota, M. Dong, H. Qi, and K. Li, “Performance guaranteed
[13] X. Zhang, J. Wang, and H. V. Poor, “Optimal resource allocation for computation offloading for mobile-edge cloud computing,” IEEE Wireless
statistical QoS provisioning in supporting mURLLC over FBC-driven Commun. Lett., vol. 6, no. 6, pp. 774–777, Dec. 2017.
6G terahertz wireless nano-networks,” in Proc. IEEE Conf. Comput. [37] G. Liu, I. Khalil, and A. Khreishah, “Using single-step adversarial training
Commun., 2021, pp. 1–10. to defend iterative adversarial examples,” in Proc. 11th ACM Conf. Data
[14] D. Xiang, “Test compression for launch-on-capture transition fault test- Appl. Secur. Privacy, 2021, pp. 17–27.
ing,” ACM Trans. Des. Automat. Electron. Syst., vol. 29, no. 1, pp. 1–20, [38] N. Zhao, Z. Ye, Y. Pei, Y. Liang, and D. Niyato, “Multi-agent deep
2023. reinforcement learning for task offloading in UAV-assisted mobile edge
[15] P. Dai, K. Hu, X. Wu, H. Xing, and Z. Yu, “Asynchronous deep reinforce- computing,” IEEE Trans. Wireless Commun., vol. 21, no. 9, pp. 6949–6960,
ment learning for data-driven task offloading in MEC-empowered vehic- Sep. 2022.
ular networks,” in Proc. IEEE Conf. Comput. Commun., 2021, pp. 1–10. [39] H. Ma, P. Huang, Z. Zhou, X. Zhang, and X. Chen, “GreenEdge: Joint
[16] M. Gu, J.-S. Yang, and J.-M. Chang, “Assessing network reliability through green energy scheduling and dynamic task offloading in multi-tier edge
perspectives of neighbor connectivity and subversion,” IEEE Trans. Netw. computing systems,” IEEE Trans. Veh. Technol., vol. 71, no. 4, pp. 4322–
Sci. Eng., vol. 11, no. 5, pp. 4384–4396, Sep./Oct. 2024. 4335, Apr. 2022.
[17] G. Wu, Y. Zhao, Y. Shen, H. Zhang, S. Shen, and S. Yu, “DRL-based [40] X. Jia and L. Zhao, “RAEF: Energy-efficient resource allocation through
resource allocation optimization for computation offloading in mobile edge energy fungibility in serverless,” in Proc. IEEE 27th Int. Conf. Parallel
computing,” in Proc. IEEE Conf. Comput. Commun. Workshops, 2022, Distrib. Syst., 2021, pp. 434–441.
pp. 1–6.
[18] J. Xu, L. Chen, and P. Zhou, “Joint service caching and task offloading for
mobile edge computing in dense networks,” in Proc. IEEE Conf. Comput.
Commun., 2018, pp. 2017–215.
[19] X. Wang, J. Ye, and J. C. Lui, “Decentralized task offloading in edge
computing: A multi-user multi-armed bandit approach,” in Proc. IEEE
Conf. Comput. Commun., 2022, pp. 1199–1208.
[20] Anon. Google.com internal network dataset connect [EB/OL]. [Online].
Available: http://connect.unikoblenz.de/networks/cfinderGoogle Fu Xiao (Senior Member, IEEE) received the PhD
[21] Anon, Notre dame network dataset konect [EB/OL]. [Online]. Available: degree in computer science and technology from
http://konect.uni-koblenz.de/networks/web-NotreDame the Nanjing University of Science and Technology,
[22] Anon. Berkeley/Stanford network dataset connect [EB/OL]. http:// Nanjing, China, in 2007. He is currently a professor
connect.uni-koblenz.de/networks/web-BerkStan and the PhD supervisor with the School of Computer,
[23] W. Chen, I. Paik, and P. Hung, “Constructing a global social service Nanjing University of Posts and Telecommunica-
network for better quality of web service discovery,” IEEE Trans. Serv. tions. He has authored papers in research related inter-
Comput., vol. 8, no. 2, pp. 284–298, Mar./Apr. 2015. national conferences, including INFOCOM, ICC, and
[24] X. Ma et al., “Optimization modeling and analysis of trustworthiness de- IPCCC. He has authored IEEE Journal on Selected
termination strategies for service discovery of MSNP,” J. Supercomputing, Areas in Communications, IEEE/ACM Transactions
vol. 75, no. 4, pp. 1766–1782, 2019. on Networking, IEEE Transactions on Dependable
[25] J. Shi, K. Fu, J. Wang, Q. Chen, D. Zeng, and M. Guo, “Adaptive and Secure Computing, IEEE Transactions on Parallel and Distributed Systems,
QoS-aware microservice deployment with excessive loads via intra-and IEEE Transactions on Mobile Computing, ACM Transactions on Embedded
inter-datacenter scheduling,” IEEE Trans. Parallel Distrib. Syst., 2024, Computing Systems, and IEEE Transactions on Vehicular Technology. His
pp. 1565–1582. research interest are the Computer networks, Internet of Things.

Authorized licensed use limited to: Indian Institute of Information Technology Design & Manufacturing. Downloaded on May 16,2025 at 04:55:41 UTC from IEEE Xplore. Restrictions apply.
 XIAO et al.: JOINT SERVICE DEPLOYMENT AND TASK OFFLOADING FOR DATACENTERS WITH EDGE HETEROGENEOUS SERVERS 853

Weibei Fan (Member, IEEE) received the PhD de- Tie Qiu (Senior Member, IEEE) is currently a full
gree in computer science from Soochow University, professor with the School of Computer Science
in 2019. He is currently an associate professor in and Technology, Tianjin University, China. He has
the school of Computer, Nanjing University of Posts authored/co-authored 10 books, more than 200 scien-
and Telecommunications. He has authored papers tific papers in international journals and conference
in research related international journals and con- proceedings, such as IEEE Transactions on Multime-
ferences, such as IEEE Transactions on Computers, dia, IEEE Transactions on Mobile Computing, IEEE
IEEE Transactions on Parallel and Distributed Sys- Transactions on Knowledge and Data Engineering,
tems, IEEE Transactions on Dependable and Secure IEEE Transactions on Industrial Informatics, IEEE
Computing, IEEE Transactions on Reliability, IEEE Transactions on Control Systems Technology, IEEE
Transactions on Network Science and Engineering, Communications, INFOCOM, GLOBECOM etc. He
IEEE Transactions on Network and Service Management, IEEE Transactions is a Senior Member of ACM. He serves as an associate editor of IEEE/ACM
on Vehicular Technology, IEEE Transactions on Emerging Topics in Compu- Transactions on Networking, IEEE Transactions on Network Science and Engi-
tational Intelligence. His research interests include data center networks and neering, and IEEE Transactions on Systems, Man, and Cybernetics.
interconnection networks.

Xiuzhen Cheng (Fellow, IEEE) received the MS and

PhD degrees in computer science from the University
of Minnesota Twin Cities, Minneapolis, Minnesota,
Lei Han is currently a professor with the School of in 2000 and 2002, respectively. She is currently a
Computer, Nanjing University of Posts and Telecom- professor with the School of Computer Science and
munications. He has created patents which are au- Technology, Shandong University, China. Her current
thorized by China, US, U.K., Japan,Korea, and Ger- research interests include privacy-aware computing,
many. He has authored papers in research related wireless and mobile security, dynamic spectrum ac-
international conferences and journals, such as IEEE cess, mobile handset networking systems (mobile
Transactions on Parallel and Distributed Systems, health and safety), cognitive radio networks, and al-
GLOBECOM, MOBISYS and ICCN. His research gorithm design and analysis. She has served on the
interest in Computer networks, Datacenter systems. editorial boards of several technical publications and the Technical Program
Committees of various professional conferences/workshops.

Authorized licensed use limited to: Indian Institute of Information Technology Design & Manufacturing. Downloaded on May 16,2025 at 04:55:41 UTC from IEEE Xplore. Restrictions apply.