Multi-Objective Optimization for Cost and Latency in Computing

Force Network

Shizhan Lan

1,2

, Siyuan Song

, Zhenyu Wang

and Yuxuan Long

2,*

China Mobile Guangxi Branch Co., Ltd, Nanning, China

South China University of Technology, Guangzhou, China

Keywords: Computing Force Network, Workflow Scheduling, Price-Sensitive.

Abstract: Computing Force Network (CFN) is a new infrastructure based on cloud, edge and end three-layer network

architecture. In CFN, tasks exist in the form of micro-services, so how to reduce the cost of user micro-

services and ensure the quality of service is a challenging problem. In order to solve the above problems,

firstly, we established the hierarchical model, resource limitation model, price model and time delay model

of micro-service workflow. Secondly, we modeled the micro-service scheduling problem under the computing

network into a multi-objective optimization problem with resource limitation as constraint and micro-service

cost and overall time delay as targets. Thirdly, we propose a multi-objective optimization algorithm based on

NSGA-II to solve the above problems. The experimental results show that the multi-objective optimization

model established in this paper is effective, and the multi-objective optimization algorithm proposed in this

paper is superior to the existing algorithms, which can effectively reduce the micro-service delay and cost.

1 INTRODUCTION

In recent years, with the development of technologies

such as 5G and the Internet of Things, the amount of

data in the global network has grown exponentially,

and the constantly surging amount of data has brought

huge challenges to cloud computing data centers.

Edge computing is an emerging distributed

computing paradigm. By extending the computing

capacity of cloud data centers to the edge of networks,

part of the data in the network can be processed by

the edge, relieving the pressure on cloud data centers

to some extent (J. Zhang, 2018). For example, from

large cloud data centers to scattered Mobile Edge

Computing (MEC) servers to mobile smart devices,

the traditional network architecture has gradually

evolved into cloud, edge, and end three-layer network

architecture. In order to make better use of

heterogeneous resources distributed in different

places and realize accurate matching between user

resource demand and heterogeneous resource supply,

Computing Force Network is proposed.

In CFN, tasks submitted by users are usually

divided into multiple micro-services and assigned to

different compute nodes for processing. The compute

nodes communicate and exchange data through

network connections to achieve collaborative

calculation and result summary (Islam A, 2021). In

order to ensure the QoS of user tasks, the following

aspects need to be considered.

Firstly, Cost is the key constraint. Due to the

heterogeneity of the underlying hardware of the

server and the interference of various factors, the

rental price and cost of different servers vary greatly.

However, users always hope that the rental cost will

not exceed the cost budget when renting servers to

deploy micro-services. Especially for high-

performance computing micro-services, AI

accelerator is expensive, and reasonable scheduling

of micro-services can save costs for enterprises.

Therefore, how to schedule micro service under strict

price cost budget is the primary consideration (Tang

X, 2022).

Secondly, Resource requirements are

heterogeneous. Micro-services in computing

networks have heterogeneous resource requirements.

In addition to general resource requirements such as

CPU and memory, many micro-services also require

more dimensions of resource allocation according to

specific business requirements. Therefore, the

heterogeneity of micro-service resource requirements

increases the complexity of micro-service scheduling.

Thirdly, the network becomes the bottleneck of

application QoS. The scheduling problem of micro-

services can be regarded as the scheduling of

Lan, S., Song, S., Wang, Z. and Long, Y.

Multi-Objective Optimization for Cost and Latency in Computing Force Network.

DOI: 10.5220/0012281700003807

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 2nd International Seminar on Artiﬁcial Intelligence, Networking and Information Technology (ANIT 2023), pages 291-300

ISBN: 978-989-758-677-4

291

workflow with dependency relationship. There is a

sequence of execution among micro-services, and

there is a relationship of data transmission between

different micro-services. In the computing power

network, computing resources are connected through

the network. Optimizing the communication delay of

micro-service is of great significance to reduce

network congestion and ensure the application of

QoS.

Our contributions are multifold and can be

summarized as follows:

Firstly, considering the cost of micro-service

scheduling, we set up a cost model from the

perspective of user micro-service scheduling cost.

Secondly, we take into account the resource

constraint of micro-service scheduling. Nodes that do

not meet the resource constraint are not allowed to

serve as scheduling nodes, so as to ensure the service

quality of micro-service, which is different from

previous workflow scheduling studies.

Then, we set up a multi-objective optimization

model with resource constraints and cost and time

delay as optimization objectives.

Finally, we propose a target capture optimization

model based on NSGA-II to solve the above

problems.

2 RELATED WORK

Many scholars have carried out in-depth research on

the optimization of pricing cost and delay of micro-

service scheduling.

According to the number of scheduling objectives,

the current micro-service scheduling can be divided

into single-objective optimization micro-service

scheduling and multi-objective optimization micro-

service scheduling. In the single-objective

optimization micro-service scheduling, only one

index is optimized, so the scheduling result is too

limited. In the micro-service scheduling with multi-

objective optimization, considering multiple

constraints and optimization objectives, the

scheduling results are more applicable. According to

the types of micro-service scheduling, micro-service

scheduling can be divided into mutually independent

micro-service scheduling and workflow scheduling.

The mutually independent micro-service scheduling

does not consider the dependency between micro-

services, while workflow scheduling considers the

execution sequence of micro-services, and its

scheduling implementation is more complex. Micro-

service scheduling algorithm can be divided into

heuristic scheduling algorithm and meta-heuristic

scheduling algorithm.

For the delay problem of microservice scheduling,

H. Topcuoglu proposed a Heterogeneous earliest-

finisher (HEFT) algorithm and a Critical-Path-on-a-

Processor, heterogeneous earlier-finisher (HEFT)

algorithm (Topcuoglu H, 2002). In the CPOP

algorithm, HEFT selects the task with the highest

ascending rank value in each step and assigns the

selected task to the processor, which minimizes its

earliest completion time using the insertion-based

method. In the CPOP algorithm, the priority of each

task is calculated by comprehensively considering the

ascending and descending sort. Since the above two

algorithms were proposed, many scholars have

proposed many improved algorithms based on the

ideas of the above two algorithms according to

different problem scenarios. Xiumin Zhou et al.

proposed a heterogeneous earliest completion time

(FDHEFT) algorithm based on fuzzy dominance

sorting, which closely combines the fuzzy dominance

sorting mechanism with the list scheduling heuristic

HEFT, while optimizing the scheduling cost and

delay (Zhou X, 2019). Faragardi et al. proposed a new

resource supply mechanism and workflow scheduling

algorithm GRP-HEFT, which is used to minimize the

maximum completion time of a given workflow, so

as to meet the budget constraints of the pay-as-the-

volume cost model in modern IaaS cloud (Faragardi

H R, 2020). In view of workflow scheduling

problems, the above algorithms optimize the delay of

workflow scheduling under the condition of

satisfying workflow cost constraints. However, the

above algorithms schedule with virtual machine as

granularity, resulting in a large amount of resource

waste. Moreover, the above algorithms do not

consider the critical path of tasks as a whole, so it is

easy to fall into local optimal. In order to implement

global scheduling of micro-service, some scholars

propose to use heuristic algorithm to solve micro-

service scheduling problem. Lin et al. proposed an ant

colony algorithm for solving scheduling problems,

which not only considered the calculation of physical

nodes and the utilization rate of storage resources, but

also the number of micro-service requests and failure

rate of physical nodes. Experimental results showed

that the algorithm achieved better results in

optimizing cluster business reliability, cluster load

balancing and network transmission overhead(Lin M,

2019). Aiming at minimizing the cost of micro-

service scheduling, Hussain et al proposed a hybrid

cuckoo search and genetic algorithm HFSGA

algorithm to realize micro-service scheduling

(Hussain S M, 2022). But their approach is also

ANIT 2023 - The International Seminar on Artiﬁcial Intelligence, Networking and Information Technology

292

virtual machine granularity, resulting in a waste of

resources. Liang et al. proposed an heuristic micro-

service scheduling algorithm based on container to

solve the scheduling problem of application

workflow based on micro-service with minimum end-

to-end delay under user-specified budget constraints

(Bao L, 2019). This algorithm optimizes the delay of

micro-service scheduling. His method takes container

as unit for scheduling, which improves resource

utilization in the scheduling process. However, it

ignores the resource limitation of nodes. Once the

resource allocation of containers exceeds the resource

capacity of nodes, QoS applied by users will be

seriously affected、. To sum up, the existing methods

have some problems, such as too large granularity of

micro-service scheduling, ignoring resource

constraints and price and cost factors, which cannot

meet the micro-service scheduling requirements

under the computing power network.

3 SOLUTIONS

In the following, we first model the micro-service

scheduling problem under the computing network,

and then propose a heuristic algorithm based on

NSGA-II to solve the above problem model.

3.1 Subsection System Model

This scenario consists of multiple physical server

servers in geographically remote locations. The nodes

are connected over the core network. Each node has

heterogeneous resources, such as CPU, memory,

bandwidth, and GPU. User tasks are broken down

into multiple micro-services with dependencies.

Micro-services exist in the form of containers. When

users use node resources, they are rented as VMS.

Users need to pay the VM rental fee. The fee is

determined by the VM rental price per unit time and

the VM rental duration. The execution of micro-

services and data transmission between micro-

services will result in a certain delay, and the overall

delay of micro-services will affect the QoS of

applications. When the micro-service is scheduled to

a node, the amount of micro-service resources

requested cannot exceed the remaining resources of

the node. Before dispatching the service, the user will

inform the cloud manufacturer of the price

expectation and hope to obtain the highest QoS within

the price expectation. The system model is shown in

figure 1.

Figure 1. System model.

Figure 2. Workflow model.

3.2 Microservice Workflow Model

The workflow between microservices can be

represented by a directed acyclic graph. Figure 2

describes the workflow structure of two

microservices with dependencies.

For any microservice workflow, it can be

expressed by 𝐺



𝑇,𝐸



, where 𝑇𝑡



,𝑡



,…,𝑡





represents a set of microservices that are dependent

on each other, 𝐸𝑒

,

|𝑡



,𝑡



∈𝑇 represents a direct

dependency between microservices, If the

microservice𝑡



depends on the microservice 𝑡



, 𝑒

,

1, otherwise 𝑒

,

is 0. Set 𝑝𝑟𝑒𝑡



 represent all

precursor nodes of microservice 𝑡



, 𝑓𝑟𝑜𝑛𝑡𝑡





represents the direct precursor of microservice 𝑡



, and

use 𝑎𝑓𝑡𝑒𝑟𝑡



 to represent the direct successor of

microservice 𝑡



In this paper, the hierarchical workflow model

mentioned in literature(

Rizvi N, 2020

) is used to

process the above workflows. The specific process is

to divide the DAG graph into multiple microservice

chains. The entry microservice of each microservice

chain has no precursor node, and the exit

microservice has no successor node. According to the

split microservice chain, the tasks are divided into

different levels, and each level contains a set of

independent microservices. For example, workflow 1

in Figure 4 can be divided into four microservice

chains: 𝑡



-𝑡



-𝑡



-𝑡



、𝑡



-𝑡



-𝑡



-𝑡



、𝑡



-𝑡



-𝑡



-𝑡



、𝑡



-𝑡



𝑡



-𝑡



. In each microservice chain, microservice levels

are divided. For example, in the 𝑡



-𝑡



-𝑡



-𝑡



chain, if

Multi-Objective Optimization for Cost and Latency in Computing Force Network

293

𝑡



is the entry node, 𝑡



is the first layer, and

accordingly 𝑡



、 𝑡



、 𝑡



are the second, third, and

fourth layers respectively. Finally, for any

microservice 𝑡



, if 𝑡



is in multiple microservice

chains, the levels of 𝑡



in each microservice chain

are𝑟𝑎𝑛𝑘



、𝑟𝑎𝑛𝑘



、…、 𝑟𝑎𝑛𝑘



, the final 𝑡



level is

the highest of the preceding levels. For example, in

workflow 2 in Figure 4, 𝑡



is in the microservice

chain 𝑡



-𝑡



-𝑡



-𝑡



、 𝑡



-𝑡



-𝑡



and 𝑡



-𝑡



-𝑡



-𝑡



. The

corresponding levels of 𝑡



in each chain are 2, 1, and

2 respectively, so the final level of 𝑡



is 2.

3.3 Resource Constraint Mode

Each VM has a certain amount of heterogeneous

resources. When microservices are scheduled to

VMS, they must meet the resource restrictions of VM

nodes. Use 𝑉𝑣



,𝑣



,…,𝑣



 to represent the set of

all virtual machine nodes. For any virtual machine

node v_j, there is a certain amount of heterogeneous

resources. This paper considers four heterogeneous

resources, namely CPU, memory, bandwidth, and

GPU. 𝑅





、 𝑅





、 𝑅





、 𝑅





are used to

represent the remaining amount of four

heterogeneous resources on the node at time t,

respectively. The application amount of

heterogeneous resources applied by microservice 𝑡



is expressed by 𝑟





、 𝑟





、 𝑟





、 𝑟





respectively. The 0-1 variable 𝑧

,

indicates whether

the microservice 𝑡



is scheduled to node 𝑣



. When

𝑧

,

is 1, it indicates that the microservice 𝑡



scheduled to 𝑣



. When 𝑧

,

is 0, it indicates that the

microservice 𝑡



is not scheduled to 𝑣



. The resource

cannot be preempted. The requested resource is

released after the microservice is executed. The

following constraints must be met during

microservice scheduling

𝑧

,







∈,

1 1

𝑟





𝑧

,

∀



∈

𝑅





2

𝑟





𝑧

,

∀



∈

𝑅





3

𝑟





𝑧

,

∀



∈

𝑅





4

𝑟





𝑧

,

∀



∈

𝑅





5

Formula (1) indicates that all microservices must

be scheduled and can only be scheduled to one node

at a time. Formulas (2) to (5) indicate that when

microservices are scheduled to any node, the number

of heterogeneous resources applied for microservices

must be less than or equal to the remaining resources

on the node.

3.4 Price-Cost Model

When a user rents a VM, the cloud vendor charges the

user a fee based on whether the user rents the VM and

the VM usage time, regardless of how many

microservice containers the user schedules on the

VM. The total price that the user needs to pay is

represented by Cost, and the calculation formula of

Cost is shown in formula (6).

𝐶𝑜𝑠𝑡 𝑝









𝑡









𝑡_𝑣



indicates the total duration of VM 𝑣



rental,

and 𝑝



indicates the unit price of VM 𝑣



rental. The

user will submit a price budget before microservice

scheduling, so the overall price cost should be lower

than the budget after the final microservice execution

is completed, otherwise the scheduling will fail. This

paper deals with the budget by first optimizing the

microservice delay and cost at the same time, finally

getting the Pareto frontier, and then calculating the

scheduling scheme with the lowest delay within the

budget according to the price budget. Therefore, the

final cost will be as close to the budget as possible, so

as to obtain the best QoS within the budget. The price

of a VM is related to the computing power of the VM

per unit computing resource. Generally, the higher the

price of a VM, the greater the computing power of the

VM per unit computing resource, and the shorter the

execution time of microservices.

3.5 Delay Model

In this paper, the time delay from the start of the first

microservice to the end of the last microservice will

be referred to as makespan, makespan is calculated as

𝑚𝑎𝑘𝑒𝑠𝑝𝑎𝑛max

∀



∈

𝐹𝑇



𝑡







Where 𝐹𝑇



𝑡





represents the total time taken from

the start of scheduling the first microservice to the

completion of the microservice 𝑡



, and formulas (7)

calculate the total time taken for all microservices to

ANIT 2023 - The International Seminar on Artiﬁcial Intelligence, Networking and Information Technology

294

be completed. For any microservice 𝑡



, 𝐹𝑇



𝑡





consists of two parts: 𝑊𝑇



𝑡





, the waiting time

required for task execution, and its own execution

time 𝑒𝑥𝑐

,

, where 𝑊𝑇



𝑡





and 𝐹𝑇



𝑡





are calculated

respectively.

𝑊𝑇



𝑡





max

∀



∈









𝐹𝑇𝑡



𝑡𝑟𝑎𝑛𝑠𝑡



,𝑡



 8

𝐹𝑇



𝑡





𝑊𝑇



𝑡





𝑒𝑥𝑐

,

9

In the above formula, assuming that the

microservice 𝑡



is at layer n, 𝑊𝑇



𝑡





represents the

total time taken for the first layer N-1 microservices

to complete. 𝑡𝑟𝑎𝑛𝑠𝑡



,𝑡



 indicates the data

transmission delay of the precursor node 𝑡



microservice 𝑡



. If the containers of microservice 𝑡



and microservice 𝑡



are scheduled to the same VM,

the delay is ignored. Otherwise, the delay is the ratio

of the size of the data transfer between the two

microservices to the average bandwidth allocated by

the container in which the two microservices reside.

Assuming that the amount of data transfer between

microservice 𝑡



and microservice 𝑡



is 𝑙𝑒𝑛𝑔𝑡ℎ

,

, the

formula for calculating 𝑡𝑟𝑎𝑛𝑠𝑡



,𝑡



 is:

𝑡𝑟𝑎𝑛𝑠𝑡



,𝑡





𝑙𝑒𝑛𝑔𝑡ℎ

,

𝑟





𝑟





/2





𝑒𝑥𝑐

,

indicates the execution delay required to

schedule microservice 𝑡



to node 𝑣



. The delay is

negatively correlated with the computing power per

unit computing resource of the VM. This paper

assumes that the delay is known.

3.6 Overall Model Design

Generally, the price of a virtual machine is related to

the computing power per unit of computing resource

of a virtual machine. The higher the price of a virtual

machine, the greater the computing power per unit of

computing resource of a virtual machine, and the

shorter the execution time of a microservice.

However, the size of virtual machine computing

power and the price of virtual machine is not a

constant proportion, under normal circumstances, the

price of virtual machine is far more than doubled

when the virtual machine computing power is

doubled. Therefore, excessive pursuit of delay

reduction will make the final price exceed the user's

cost budget. Similarly, if only lower cost is required,

the delay of the entire application will increase,

affecting the QoS of the application. Therefore, the

optimization direction of price cost and delay is not

consistent, so that both objectives can be optimized,

so that a relatively optimal scheduling scheme can be

obtained under each price budget. Combined with the

above problem description and the general model

formula of multi-objective optimization introduced in

Section 4.2, the above problem is modeled into a

multi-objective optimization model in this paper, as

shown below.

The objective function is:

𝐶𝑜𝑠𝑡𝑀𝑖𝑛 𝑝









𝑡





11

𝑚𝑎𝑘𝑒𝑠𝑝𝑎𝑛𝑀𝑖𝑛 max

∀



∈

𝐹𝑇



𝑡





The relevant constraints are:

𝑧

,



∀



∈,

1

𝑟





𝑧

,

∀



∈

 𝑅





𝑟





𝑧

,

∀



∈

 𝑅





𝑟





𝑧

,

∀



∈

 𝑅





𝑟





𝑧

,

∀



∈

 𝑅





In this paper, the process of solving the final

scheduling scheme is divided into two steps: the first

step is to obtain a set of uniformly distributed feasible

solutions by solving the above multi-objective

optimization model, that is, Pareto optimal front; The

second step is to solve the optimal scheduling scheme

according to the price budget set by the user.

3.7 Muti-Objective Optimization

Algorithm Based on NSGA-II

Figure 3 shows the flow chart of the algorithm. The

input of the algorithm is microservice set, virtual

machine node set, and user price expectation, and the

output of the algorithm is Pareto optimal frontier.

In NSGA-II algorithm, the common encoding

methods include binary encoding, symbol encoding

and real encoding. The traditional binary coding and

decoding process is more troublesome, but the real

coding reduces the complexity of calculation and

improves the efficiency of operation. The goal of this

paper is to schedule m microservice containers to be

scheduled on n virtual machine nodes. Based on the

characteristics of the problems studied in this paper,

the real coding mode is selected, as shown in Figure

4. The numbers 1-9 represent the microservice to be

scheduled, Node1 to Node5 represent the number of

the VM node that can be scheduled, and the number

corresponding to the server node number indicates

Multi-Objective Optimization for Cost and Latency in Computing Force Network

295

that the microservice is scheduled to the

corresponding VM node. For example,

((1,4),(6,7),(2,3,9),(5,8)) indicates a possible initial

solution. Resource constraints must be satisfied when

generating the initial feasible solution. In order to

generate the initial population, this paper randomly

generates x initial solutions and the initial population

𝑃𝑆



,𝑆



,...,𝑆



.

Figure 3. NSGA-II Algorithm Flowchart.

Figure 4. Coding Scheme.

The initial population was sorted according to the

fitness function. The input of fast non-dominated

sequencing was the original population P and the

output was the stratified population 𝑃



For the parent population 𝑃𝑆



,𝑆



,...,𝑆



. For

any individual 𝑆



, calculate the values of objective

function 1 and objective function 2 of S_i. If for any

other individual 𝑆



in the population, 𝑆



does not have

a pareto dominance over 𝑆



then divide 𝑆



into the

first non-dominated layer and traverse the population

successively to find all the individuals meeting the

above conditions. Divide all of the above individuals

into the current tier and delete all of the above

individuals into the current tier from the original

population.

The number of non-dominant layers of the

population is increased by one each time, and the

above steps are repeated until there are no individuals

in the original population, and finally the stratified

population 𝑃



𝑆



,𝑆



,...,𝑆



 .The number of

non-dominated levels of an individual represents the

quality of the solution, and the smaller the number of

non-dominated levels, the better the performance of

the individual and the closer to the optimization goal.

The crowding degree 𝑖



of each individual in each

layer of the stratified population 𝑃



was calculated in

turn. The degree of crowding represents the density

of individuals around an individual in the population,

and the value is equal to the circumference of the

rectangle with the vertex near the point. Let the

crowding degree of individuals 𝑂



and 𝐼



at the

boundary position be ∞, and the formula for

calculating the crowding degree of individuals at the

other positions be

𝑖





│𝑓





𝑓





│

│𝑓





𝑓





│











m is the number of fitness evaluation functions,

𝑓





and 𝑓





represent the function value of the JTH

objective of the i+1 individual and the I-1 individual,

respectively, 𝑓





and 𝑓





represent the maximum

and minimum objective function values of all

individuals in the current level for the objective j,

respectively.

The elite selection strategy is based on non-

dominant ordering and crowding distance to obtain

progeny populations. Suppose that for the stratified

population 𝑃



𝑆



,𝑆



,...,𝑆



, each layer is

sorted in ascending order by crowding distance, and

the steps selected by the elite are: All the individuals

from the first layer in 𝑃



were added to the new

population P, and then all the individuals from the

second layer were added to the new population P, and

so on, until the individuals from a certain layer could

not all be added to the new population P, and the

individuals from that layer were added to P in the

order of the crowding degree distance, until the

ANIT 2023 - The International Seminar on Artiﬁcial Intelligence, Networking and Information Technology

296

number of individuals in the new population P

reached x, as shown in Figure 5.

Figure 5. Fast non-dominated sorting algorithm

Figure 6. Workflow

4 EXPERIMENT

Table 1. Specifications and Prices of the Six VMS.

Instance CPU Memory GPU Bandwidth cost

𝑣



4 16 4 100 30.39

𝑣



4 16 8 100 60.94

𝑣



8 32 16 100 130.88

𝑣



6 32 24 100 150.09

𝑣



8 40 32 100 180.04

𝑣



16 40 32 100 190.10

Table 2. NSGA-II Parameter Setting.

Parameter number

Size 50

Number of iterations 50

Cross probability 1

Mutation probability 0.1

In this section, the NSGA-II-based microservice

scheduling algorithm is tested using the Cloudsim

simulation platform, which has the modeling and

simulation functions of physical machines and

containers. This paper first analyzes several common

workflow structures in Alibaba Cluster Trace

Program, and constructs DAG graphs with 5, 10, 15

and 20 microservices respectively by referring to

common workflow structures. Figure 6 shows the

DAG diagram when the number of microservices is

10, and the weights on the edges of the diagram

represent the size of the data transfer volume of the

microservices with dependencies.

The specifications and prices of the VMS used in

the experiment refer to the cloud vendor's charging by

volume rules. Table 1 lists the specifications and

prices of the six VMS used in this paper.

The parameter Settings of an algorithm largely

determine the performance of the algorithm. Table 2

lists the parameters of the algorithm in this paper.

Price and microservice delay are used as

evaluation indexes for microservice scheduling. The

calculation formulas for the above two are formula

(6) and formula (7) respectively. To verify the

performance of the scheduling algorithms in this

paper, the Spread, Binpack, and HEFT algorithms are

selected as benchmarks. Spread and Binpack

algorithms are common methods in container

scheduling. Spread tends to distribute containers to

each node to balance cluster load, while Binpack

tends to dispatch containers to one node to improve

resource utilization. HEFT algorithm is a classic

algorithm in workflow scheduling. Its idea is to

always schedule tasks to the node with the minimum

completion time. However, HEFT algorithm

schedules tasks based on virtual machines and does

not consider resource constraints during scheduling.

The HEFT is changed to a HEFT algorithm that

schedules by container and considers resource

constraints.

Figure 7 shows the scheduling success rates of the

four algorithms at different price expectation levels.

Subgraphs (a), (b), (c) and (d) respectively show the

scheduling success rates of four algorithms with DAG

sizes of 5, 10, 15 and 20. The higher the price

expectation level, the more adequate the price budget

given by the user. Because the Spread scheme tends

to schedule microservices to different nodes, a large

number of virtual machines are rented, and the data

transmission delay between microservices becomes

longer, which ultimately makes scheduling

impossible under the condition of meeting the price

constraint. The Binpack scheme tends to schedule

microservices to a node, so the number of leased

virtual machines is small and the communication

delay between microservices is reduced, which can

meet the price constraint to a certain extent, but it

Multi-Objective Optimization for Cost and Latency in Computing Force Network

297

(a) 5 microservices

(b) 10 microservices

(d) 20 microservices

Figure 7. Success Rate.

cannot take into account the global scheduling, so it

cannot meet the scheduling demand when the price

constraint level is high. The HEFT algorithm always

schedules microservices to the node with the shortest

completion time, without considering the global

scheduling and scheduling cost, so the scheduling

result is difficult to meet the cost expectation set by

users. The NSGA-II-based microservice scheduling

algorithm proposed in this paper also optimizes the

scheduling delay and cost of microservices, so the

scheduling success rate of the algorithm proposed in

this paper exceeds other algorithms.

Figure 7 shows the scheduling delay of four

microservice algorithms under four microservice

scales and different price expectations. Subfigures

(a), (b), (c), and (d) show the experimental results

when DAG scales are 5, 10, 15, and 20 respectively.

It can be seen from the figure that with the increase of

price expectation, the delay of various algorithms

shows a non-increasing trend. Among them, the

algorithm proposed in this paper can obtain lower

microservice delay compared with other algorithms

under the same price expectation. When the price

expectation increases to a certain value, the delay will

no longer decrease, and higher QoS can no longer be

obtained when the price expectation is increased. The

above phenomenon is in line with normal logic,

because microservice execution and data

transmission will certainly cost a certain delay, and

the computing power of virtual machine nodes and

the transmission capacity of the network are limited,

so the delay can not be reduced. From the

experimental results, it can be seen that the algorithm

proposed in this paper can find the global relative

optimal scheduling scheme under the user-set price

expectation. For example, under 5 microservices,

when the user price expectation is 2, only the

algorithm in this paper and the Binpack algorithm can

give the scheduling scheme under the price

expectation, and the other two algorithms fail to

schedule. When the user's price expectation is 2.6, the

algorithm in this paper obtains a lower delay than

Binpack and Spread, so the user pays the same price,

and the algorithm in this paper can obtain higher QoS.

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

1234

sucess rate

Price expectation level

NSGA-II HEFT Spread Binpack

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

1234

sucess rate

Price expectation level

NSGA-II HEFT Spread Binpack

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

1234

sucess rate

Price expectation level

NSGA-II HEFT Spread Binpack

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

1234

sucess rate

Price expectation level

NSGA-II HEFT Spread Binpack

ANIT 2023 - The International Seminar on Artiﬁcial Intelligence, Networking and Information Technology

298

(a) 5 microservices

(b) 10 microservices

(d) 20 microservices

Figure 8. Scheduling delay in different DAGs.

In summary, the multi-objective optimization

algorithm proposed in this paper can optimize both

delay and price, and can select the relatively optimal

scheduling scheme from the Pareto frontier solution

according to the price expectation set by users, which

improves the scheduling success rate and reduces the

microservice delay. The algorithm proposed in this

paper provides a solution for the price-sensitive

microservice scheduling under the computing power

network.

5 CONCLUSION

This paper first analyzes the microservice scheduling

problem under the CFN, then we introduce the

relevant theories and technologies of multi-objective

optimization, and models the scheduling problem of

microservice under the computing network into a

multi-objective optimization problem. Finally, a

multi-objective optimization algorithm based on

NSGA-II is proposed to solve the above problem

model. The experimental results show that the

proposed algorithm can optimize both the price cost

and the microservice delay, and finally give a

relatively optimal scheduling scheme according to the

price expectation set by the user.

The microservice scheduling model constructed

in this paper does not take into account the oversold

problem of resources, that is, the amount of resource

applications of containers on a virtual machine node

can be greater than the total amount of resources

owned by the virtual machine. Further research can

be carried out in the future.

REFERENCES

J. Zhang, B. Chen, Y. Zhao, X. Cheng and F. Hu, "Data

Security and Privacy-Preserving in Edge Computing

Paradigm: Survey and Open Issues," in IEEE Access,

vol. 6, pp. 18209-18237, 2018,

https://doi.org/10.1109/ACCESS.2018.2820162

Islam A, Debnath A, Ghose M, et al. A survey on task

offloading in multi-access edge computing (J). Journal

of Systems Architecture, 2021, 118: 102225.

https://doi.org/10.1016/j.sysarc.2021.102225

Tang X, Cao W, Tang H. Cost-Efficient Workflow

Scheduling Algorithm for Applications With Deadline

Constraint on Heterogeneous Clouds (J). IEEE

Transactions on Parallel and Distributed Systems,

2022, 33(9): 2079–2092.

https://doi.org/10.1109/TPDS.2021.3134247

Topcuoglu H, Hariri S, Society I C. Performance-effective

and low-complexity task scheduling for heterogeneous

100

1,4 1,6 1,8 2 2,2 2,4 2,6 2,8 3 3,2 3,4 3,6 3,8

delay/s

cost expectation

NSGA-II

Spread

Binpack

HEFT

100

110

120

130

140

2,4 2,6 2,8 3 3,2 3,4 3,6 4 4,2 4,4 4,6 4,8 5

delay/s

cost expectation

100

3,6 3,8 4 4,2 4,4 4,6 4,8 5 5,2 5,4 5,6 5,8 6

delay/s

cost expectation

120

130

140

150

160

170

180

6 6,4 6,8 7,2 7,6 8 8,4 8,8 9,2 9,6 10 10,4 10,8

delay/s

cost expectation

Multi-Objective Optimization for Cost and Latency in Computing Force Network

299

computing (J). IEEE Transactions on Parallel and

Distributed Systems, 2002, 13(3): 260–274.

https://doi.org/10.1109/71.993206

Zhou X, Zhang G, Sun J. Minimizing cost and makespan

for workflow scheduling in cloud using fuzzy

dominance sort based HEFT (J). Future Generation

Computer Systems, 2019, 93: 278–289.

https://doi.org/10.1016/j.future.2018.10.046

Faragardi H R, Reza M, Sedghpour S. GRP-HEFT : A

Budget-Constrained Resource Provisioning Scheme for

Workflow Scheduling in IaaS Clouds (J). IEEE

Transactions on Parallel and Distributed Systems,

2020, 31(6): 1239–1254.

https://doi.org/10.1109/TPDS.2019.2961098

Lin M, Xi J, Bai W. Ant Colony Algorithm for Multi-

Objective Optimization of Container-Based

Microservice Scheduling in Cloud (J). IEEE Access,

2019, 7: 83088–83100. https://doi.org/10.1109

/ACCESS.2019.2924414

Hussain S M, Begh G R. Hybrid heuristic algorithm for

cost-efficient QoS aware task scheduling in fog–cloud

environment (J). Journal of Computational Science,

Elsevier B.V., 2022, 64(January): 101828.

https://doi.org/10.1016/j.jocs.2022.101828

Bao L, Wu C, Bu X, et al. Performance Modeling and

Workflow Scheduling of Microservice-Based

Applications in Clouds(J). IEEE Transactions on

Parallel and Distributed Systems, IEEE, 2019, 30(9):

2101–2116.

https://doi.org/10.1109/TPDS.2019.2901467

Rizvi N, Ramesh D. HBDCWS: heuristic-based budget and

deadline constrained workflow scheduling approach for

heterogeneous clouds (J). Soft Computing, Springer

Berlin Heidelberg, 2020, 24(24): 18971–18990.

https://doi.org/10.1007/s00500-020-05127-9

ANIT 2023 - The International Seminar on Artiﬁcial Intelligence, Networking and Information Technology

300