1074 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 31, NO.

5, MAY 2020

Customer Perceived Value- and Risk-Aware

Multiserver Configuration for Profit Maximization
Tian Wang , Student Member, IEEE, Junlong Zhou , Member, IEEE,
Gongxuan Zhang, Senior Member, IEEE, Tongquan Wei , Senior Member, IEEE,
and Shiyan Hu, Senior Member, IEEE

Abstract—Along with the wide deployment of infrastructures and the rapid development of virtualization techniques in cloud
computing, more and more enterprises begin to adopt cloud services, inspiring the emergence of various cloud service providers.
The goal of cloud service providers is to pursue profit maximization. To achieve this goal, cloud service providers need to have a good
understanding of the economics of cloud computing. However, the existing pricing strategies rarely consider the interaction between
user requests for services and the cloud service provider and hence cannot accurately reflect the supply and demand law of the cloud
service market. In addition, few previous pricing strategies take into account the risk involved in the pricing contract. In this article, we
first propose a dynamic pricing strategy that is developed based on the customer perceived value (CPV) and is able to accurately
capture the real situation of supply and demand in marketing. The strategy is utilized to estimate the user’s demand for cloud services.
We then design a profit maximization scheme that is developed based on the CPV-aware dynamic pricing strategy and considers the
risk in the pricing contract. The scheme is utilized to derive the optimal multiserver configuration for maximizing the profit. Extensive
simulations are carried out to verify the proposed customer perceived value and risk-aware profit maximization scheme. As compared
to two state of the art benchmarking methods, the proposed scheme gains 31.6 and 30.8 percent more profit on average, respectively.

Index Terms—Cloud computing, customer perceived value, dynamic pricing, multiserver configuration, profit maximization, risk

1 INTRODUCTION
computing, as a new and effective commercial the infrastructure providers. Either increasing the revenue
C LOUD
model that turns the delivery of storage, computing, and
communication resources into ordinary commodities in a
or reducing the cost can improve the profit. Thus, cloud
service providers develop a series of approaches [4], [5], [6],
pay-as-you-go manner, is being paid more and more atten- [7] to raise the profit in two aspects: one is to raise the reve-
tion by academia and industry [1]. With the rapid develop- nue by setting a high selling price for cloud services while
ment of virtualization techniques and the wide deployment attracting a large number of customers to buy services and
of infrastructures in cloud computing, the number and the other is to reduce operating cost for providing services.
variety of cloud service providers (e.g., Amazon EC2 [2] and Note that a service’s price as well as purchase amount inter-
Microsoft Azure [3]) have dramatically increased. The goal play and thus cannot be optimized simultaneously [8]. In
of these cloud service providers is to pursue profit maximi- this paper, we propose a dynamic pricing strategy which
zation. Therefore, for these service providers, how to deter- takes into account the interaction between a service’s price
mine their cloud service prices and configure their cloud and purchase amount.
servers for obtaining the maximum profit are of utmost Unlike static pricing strategies (e.g., tiered pricing, sub-
importance. scription based pricing, and pay-as-you-go used in real-
Like all the traditional businesses, the profit of a cloud ser- world cloud service providers Amazon Web Services [9] and
vice provider is the difference between the revenue gained Google App Engine [10]) that fix the price of a service request
by selling cloud services to customers and the monetary cost in advance as well as do not change the price with market
of rental charge as well as electricity bill of servers owned by conditions [5], [7], [11], dynamic pricing strategies adjust the
service price according to customer demands. Since dynamic

T. Wang, J. Zhou, and G. Zhang are with the School of Computer Science pricing strategies can better cope with unpredictable cus-
and Engineering, Nanjing University of Science and Technology, Nanjing tomer demands, they have been adopted by more and more
210094, China. E-mail: {eclipse_wt, jlzhou, gongxuan}@njust.edu.cn. cloud service providers. Amazon has introduced a spot pric-

T. Wei is with the School of Computer Science and Technology, East China
Normal University, Shanghai 200062, China. E-mail: [email protected].
ing strategy [12] to adjust the price for a virtual instance at

S. Hu is with the School of Electronics and Computer Science, University runtime. Based on an empirical study of Amazon’s spot price
of Southampton, SO17 1BJ, Southampton, United Kingdom. history, Xu and Li [13] proposed a market-driven dynamic
E-mail: [email protected]. pricing mechanism to have a better understanding of the cur-
Manuscript received 3 Apr. 2019; revised 10 Nov. 2019; accepted 11 Nov. rent market demand. A genetic algorithm based dynamic
2019. Date of publication 16 Dec. 2019; date of current version 20 Jan. 2020. pricing strategy is developed by Macias et al. [14] to find the
(Corresponding authors: Junlong Zhou and Gongxuan Zhang.)
Recommended for acceptance by J. Zhai. optimal price for maximizing profit. Ren et al. [15] presented
Digital Object Identifier no. 10.1109/TPDS.2019.2960024 a dynamic scheduling and pricing strategy for delay-tolerant
1045-9219 ß 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See ht_tp://www.ieee.org/publications_standards/publications/rights/index.html for more information.
WANG ET AL.: CUSTOMER PERCEIVED VALUE- AND RISK-AWARE MULTISERVER CONFIGURATION FOR PROFIT MAXIMIZATION 1075

batch services to maximize long-term profit for cloud service The results demonstrate that the proposed scheme not
providers. A dynamic and optimal pricing scheme is espe- only follows the market supply and demand law but also
cially designed for provisioning Sensors-as-a-Service [16]. outperforms the state of the art benchmarking approaches.
However, all the above-mentioned dynamic pricing strate- Specifically, the proposed scheme can achieve 31.6 and
gies do not consider customer perceived value, which is the 30.8 percent more profit on average compared to two bench-
fundamental of marketing activities. marking approaches, respectively.
The concept of customer perceived value (CPV) is first The rest of the paper is organized as follows. Section 2
introduced by Zaithaml et al. [17] and has been widely used introduces the system architecture and models. Section 3
in the modeling of market activities [18], [19], [20]. CPV is presents the proposed dynamic pricing strategy and Section 4
described as the customer’s overall assessment of the utility describes the proposed profit optimization scheme. The
of a product or service based on perceptions of what is proposed scheme is validated in Section 5 and concluding
received and what is given [17]. It reflects the worth of a prod- remarks are given in Section 6.
uct in the customer’s mind, and increases when customers
believe the benefits outweigh the costs. Since customers are
generally unaware of the true cost of production for the prod- 2 SYSTEM ARCHITECTURE AND MODEL
uct they buy, they will only pay for the product when their This section describes the cloud service provision structure
CPV for the product is higher than the product’s selling price. and related models used in the paper.
In such way, CPV impacts the purchase amount of services
and ultimately influences the cloud service provider’s profit.
2.1 Cloud System Architecture
In this paper, we design a CPV-aware dynamic pricing strat-
Consider a three-tier cloud service provision structure con-
egy for service providers.
sisting of three typical entities, i.e., cloud infrastructure pro-
A pricing contract in cloud computing includes many
viders, cloud service providers, and cloud customers. The
considerations, such as the user demand for a service, the
three-tier structure has been widely adopted in the literature
configuration of a multiserver system, the service-level
[1], [4], [6]. In the market formed by the three entities, the
agreement, the instructions of tasks to complete the ser-
vice, the task waiting time, and etc. Due to the inherent cloud infrastructure provider owns the hardware and soft-
uncertainty in task executions, the estimated time and ware facilities, and charges the cloud service provider for
monetary cost may be different from the actual ones, lead- renting the infrastructure resources. The cloud service pro-
ing to a decreased profit or even a loss. Therefore, cloud vider utilizes these resources to prepare services in the form
service providers should consider the risk induced by the of virtual machines (VMs), and charges customers for proc-
uncertainty in the pricing contract. In this paper, we study essing their service requests. The cloud customer submits
service requests to the service provider and pays for the serv-
the problem of maximizing the profit of cloud service
ices based on the services’ amount and quality.
providers. To solve the problem, we present a CPV-based
dynamic pricing strategy which conforms to the supply
and demand law in economics and a risk-aware profit 2.2 Cloud Service Provider Model
maximization scheme that takes into account the risk in Cloud service providers offer many services to their custom-
the pricing contract. The major contributions of this paper ers, e.g., infrastructure-as-a-service (IaaS), platform-as-a-ser-
are summarized below. vice (PaaS), and software-as-a-service (SaaS). We consider
IaaS cloud service providers in this paper. Examples of IaaS

We propose a CPV-based dynamic pricing strategy providers include Google Compute Engine [21], Microsoft
which takes into account the inter-relationship Azure [3], and Amazon CloudFormation [22]. In IaaS, com-
between the customers and providers of cloud serv- puting resources are offered as VMs to support the end users
ices. Based on the concept of CPV, the strategy devel- operations and the cloud service provider is responsible for
ops a calculation formula and a standard bisection running and maintaining the services. Owing to the cloud vir-
based heuristic to derive customers’ demand for cloud tualization techniques, the cloud service providers could
services. The correctness of the proposed pricing strat- focus on their own business without spending energy on the
egy and calculation formula are also validated. construction and maintenance of hardware platforms. We

We design a risk-aware profit maximization scheme assume that an IaaS cloud service provider solves customers’
that optimizes the cloud service provider’s profit by service requests by renting a multiserver system, which is
configuring the multiserver system. The scheme is constructed and maintained by an infrastructure vendor and
developed based on our dynamic pricing strategy has quite flexible architecture details for deploying pricing
and considers the risk in the pricing contract. It also strategy [5]. Such a multiserver system can be in the multiple
provides a simulated-annealing based heuristic to forms of infrastructures, e.g., blade centers where each server
find the numerical optimal solution to the profit maxi- is a server blade [23], clusters of traditional servers where
mization problem. each server is an ordinary processor [24], and multicore server

We conduct extensive simulation experiments to vali- processors where each server is a single core [25]. These
date the proposed scheme by i) observing the chang- blades/processors/cores are simply called servers in this
ing trend of profit in different conditions as well as paper for easy presentation. The serving process operates as
analyzing the factors affecting the profit, and ii) com- follows. Customers submit their service requests to the service
paring the performance of the proposed scheme with provider, who then runs tasks/applications on the multi-
two benchmarking approaches in increasing profit. server system to provide services for serving these requests.
1076 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 31, NO. 5, MAY 2020

Consider a multiserver system composed of M homoge- where P0 is the probability of a queue with no tasks and is
neous servers running at the same speed of s. The homoge- expressed as
neous multiserver model is simple yet effective and has been !1
XM1 ðMrÞk ðMrÞM 1
widely used in the literature [1], [4], [5], [6], [20], [25]. The P0 ¼ þ  : (3)
multiserver is typically modeled as an M/M/M queuing sys- k¼0 k! M! 1r
tem in which service requests arrive the multiserver system
with a rate of , wait in a first-come-first-served (FCFS) Let Pb denote the probability that a newly submitted service
queue with an infinite capacity when all the M servers are request waits in the queue when all the servers are busy,
busy, and get processed with a speed s when the servers are and it is formulated as
available. The arrival of service requests is governed by a X
1
PM
Poisson process and the M servers process these requests in Pb ¼ Pk ¼ ; (4)
k¼M
1r
parallel. Note that our homogeneous multiserver model can
be extended to heterogeneous multiserver systems. Like the
where PM is the probability of having M service requests in
method proposed in [26], we can configure a heterogeneous
the system and it can be obtained using Eq. (2).
cloud computing platform as multiple homogeneous multi-
For a newly arrived service request, let W and fW ðtÞ be
server systems. Each multiserver system is deployed with
the service’s waiting time and corresponding probability
special software and is devoted to serve one type of service
density function (PDF) [5], respectively. Then fW ðtÞ is
requests and applications. In a multiserver system, the serv-
ers are homogenous and execute at the same speed. Each fW ðtÞ ¼ ð1  Pb ÞuðtÞ þ MmPM eð1rÞMmt ; (5)
multiserver system can be treated as an M/M/M queuing
system which handles the requests following the FCFS disci- where uðtÞ is a unit impulse function and is expressed as
pline. Using the configuration method [26], our cloud service (
provider model for homogeneous servers can be readily z; 0 4 t 4 1z
applied to heterogeneous cloud computing systems. Note uz ðtÞ ¼ ; (6)
0; t > 1z
that the configuration method is suitable for the current real-
world cloud service providers such as Amazon EC2 [2], and uðtÞ ¼ limz!1 uz ðtÞ holds. Note that uz ðtÞ is in fact a PDF
Microsoft Azure [3], and Aliyun [27] since they mainly focus of an arbitrary random variable, thus can be treated as the
on the multiserver system’s heterogeneity from few aspects PDF of random variable r in the following [5]. The function
(e.g., CPU and memory). But eventually the heterogeneity uz ðtÞ satisfies two properties, i.e.,
could come from various aspects such as CPU cores/ Z 1
frequency, memory type/capacity, disk type/capacity, net- uz ðtÞ dt ¼ 1 (7)
work type/topology, software configurations, and etc. The 0
configuration method then becomes practically infeasible in Z 1 Z 1=z
1
this case since it may need infinite combinations of homoge- tuz ðtÞ dt ¼ z t dt ¼ : (8)
0 0 2z
neous multiserver systems. Thus, the scalability of the exten-
sion method should be improved for coping with all the Given these, the cumulative distribution function of W is
possible scenarios. We leave the detailed discussion of this PM Mmð1rÞt
derived as FW ðtÞ ¼ 1  1r e .
aspect to future work due to page limit.
For a customer’s request in the M/M/M queuing sys- 2.3 Service Level Agreement Model
tem,1 the service execution time is an exponential random
To ensure the customer satisfaction, we need to stipulate a
variable, represented by x ¼ r=s with mean x ¼ r=s, where
service-level agreement (SLA) between service providers and
r is the number of instructions to be executed for completing
customers on the price and the service quality. A widely-
the service request. Let m denote the service rate of customer used SLA model [30] defining the service charge for a service
requests, which is in fact the average number of service request with execution requirement r and response time R is
requests completed by a server with speed s in one unit of adopted in the paper. That is
time. The rate m is formulated as m ¼ 1=x ¼ s=r. Let r be the
8
utilization of a server, which is defined as the average per- > ar; if 0 4 R 4 scr0
<
centage of time that the server is busy and is computed as Cðr; RÞ ¼ ar  gðR  scr0 Þ; if scr0 < R < ðga þ sc0 Þr;
>
:
  r 0; if R 5 ðga þ sc0 Þr
r¼ ¼ ¼ : (1)
Mm Mðs=rÞ Ms (9)
where a is the service charge per unit amount of service as
Suppose Pk is the probability that k service requests are well as g is a coefficient indicating the compensation strength
waiting or processing in the M/M/M queuing system. because of low service quality. c is a constant related to the
According to the queuing theory [29], Pk is derived as SLA and s0 is the expected service processing speed of cus-
8 tomers. From Eq. (7), it is clear that the service is charged
< P ðMrÞk ; k 4 M based on the service response time R. Specifically, the service
0 k!
Pk ¼ ; (2)
: P M M rk ; k > M charge function is defined as follows. (i) If the response time
0 M! R is shorter than scr0 , indicating a high quality of service is pro-
vided since the service request is processed at the expected
1. Simple examples of the queuing system can be found in [28]. speed of customers, then the service charge to a customer is
WANG ET AL.: CUSTOMER PERCEIVED VALUE- AND RISK-AWARE MULTISERVER CONFIGURATION FOR PROFIT MAXIMIZATION 1077

ar which is linearly proportional to the task execution The PoS satisfaction of a user for a service, UPoS , defined as
requirement. (ii) If the response time R is between the inter- the comparison between the expected price and the actual
val ½cr=s0 ; ða=g þ c=s0 Þr, indicating a low quality of service price, is formulated as [31]
is provided since the processing time of the service request
exceeds the time at the expected speed, then the charge UPoS ¼ eðCexp Cact Þ=Cexp ; (13)
decreases as the response time R increases. (iii) If R is longer
than ðga þ sc Þr, then the service is provided as free since the
0 where Cexp and Cact represent user’s expected price and serv-
request is waiting too long in the queue.
ice’s actual price, respectively. Clearly, if the actual price Cact
The response time R of a service request is calculated as
is equal to the expected price Cexp , the default PoS satisfaction
the sum of the waiting time W and the processing time r=s,
UPoS is set to 1, indicating that the price doesnot impact the
that is, R ¼ W þ r=s. Substituting R ¼ W þ r=s into Eq. (7),
total satisfaction. If Cact is higher than Cexp , the user would be
the service charge function can be rewritten as
disappointed by the high price, thus UPoS is less than 1 and
8 decreases with the increasing actual price. If Cact is lower
> ar; if 0 4 W 4 ðsc0  1sÞr
>
> than Cexp , the user would be delighted by the low price, thus
< ða þ cg  g Þr  gW; if ð c  1Þr < W
s0 s s0 s
Cðr; W Þ ¼ : UPoS is greater than 1 and increases as the actual price
>
> 4 ðga þ sc0  1sÞr
>
:
decreases. The PoS (i.e., the charge of service) is determined
0; if W > ðga þ sc0  1sÞr by W and r. As the three modes of PoS defined in Eq. (8), the
(10) PoS satisfaction UPoS needs to consider the three cases. (i)
Cexp ¼ Cact ¼ ar, (ii) Cact ¼ ða þ cg s0  s Þr  gW and Cexp ¼ ar,
g
Since the PDF of r and W are given in Eqs. (5) and (6), the and (iii) Cact ¼ 0 and Cexp ¼ ar. Substituting the values of
expectation of the service charge can be readily derived. As
Cact and Cexp into Eq. (11), UPoS is then formulated as
shown in Eq. (8), the charge to a service request is decided
8  
by parameters a, g, c, r, W , s. In this paper, we focus on >
>
> 1; if 0 4 W 4 c
 1
s r;
homogeneous multiserver systems and thus do not con- >
>  
s0
>
>  
sider the server heterogeneity in the pricing model (i.e., > g 1 W c
< ea sþ r s0 ; if
s0  s r < W
c 1
charge function). However, our model can be extended to UPoS ðr; W Þ ¼   :
>
>
heterogeneous multiserver systems using the method [26]. >
> 4 ga þ sc0  1s
>
>  
The architecture of heterogeneous multiserver systems pre- >
>
sented in [26] indicates that the parameters considered in : e; if W > ga þ sc  1s r
0
our pricing model are sufficient but need different settings (14)
to capture the server heterogeneity. For example, our pric-
ing model should allow the instruction related parameters The user satisfaction Usat of a service is defined as the
of a service request to be changed with the service type product of QoS satisfaction and PoS satisfaction, i.e.,
and the server speed and size to be changed with the
server type. Usat ¼ UQoS ðr; W Þ  UPoS ðr; W Þ ¼
8  
The user satisfaction of a service is evaluated from two
>
> 1; if 0 4 W 4 sc0  1s r
aspects: quality of service (QoS) and price of service (PoS). >
>
>
>    
QoS is a subjective concept that is introduced to describe the >
>  
> 1þga 1ssc0 þ gac=s011=s Wr
>
discrepancy between users’ expectations on how the service <e ; if c
s0  1
s r < W
request should be served and users’ perceptions on how   ;
>
> 4 ga þ sc0  1s
the service is actually performed. The expectation is built on >
>
>
>
>
> 2 W 
the established price. Obviously, the higher the PoS is, the >
>  
higher the expectation would be. For a given established : c 1
e s0 s r ; if W > ga þ sc  1s r
0
price, if the perception (actual) performance surpasses the
expectation performance, the QoS is then deemed as high, (15)
and vice verse. Given above, the user’s QoS satisfaction for a
service, UQoS , is formulated as [31] where UQoS ðr; W Þ and UPoS ðr; W Þ are given in Eqs. (9) and
(12), respectively.

1; if Pact 5Pexp
UQoS ¼ ; (11)
e j ðPact Pexp Þ=Pexp j ; if Pact < Pexp
3 DYNAMIC PRICING STRATEGY
This section introduces the proposed dynamic pricing strat-
where Pact is the actual performance and holds for Pact ¼ W , egy for cloud service providers in detail. Since the strategy is
and Pexp is the expectation performance and holds for designed based on CPV, in this section we first describe how
Pexp ¼ ðsc0  1sÞr. Substituting Pact ¼ W and Pexp ¼ ðsc0  1sÞr to use the SERVQUAL score to model the CPV and hence
into Eq. (9), UQoS is then re-written as estimate the user’s demand for services, followed by a vali-
8 dation of the correctness of the SERVQUAL model and the
< 1; if 0 4 W 4 ðsc0  1sÞr proposed user demand function. Considering that using the
UQoS ðr; W Þ ¼ ðc=s0 1=sÞrW : analytical method to directly compute user service demand
:e ðc=s0 1=sÞr ; if W > ðsc0  1sÞr
is challenging, we also provide a heuristic algorithm to
(12) obtain user demand based on CPV.
1078 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 31, NO. 5, MAY 2020

3.1 SERVQUAL Score to Model CPV SQSwaitingtime ðrÞ ¼ SQSwaitingtime ðr; W Þ

SERVQUAL [32] is a research instrument designed for cap- !
turing consumer’s expectations and perceptions of a service c 1
¼ Pecwaitingtime  Expwaitingtime ¼ W   r
from multiple dimensions that represent service quality. s0 s
The fundamental of SERVQUAL is the gap model [33] that Z ! !
1
c 1
service quality is determined by the difference between the ¼ t  r fW ðtÞ dt
level of service perceived by users and the level of service 0 s0 s
Z ! Z
expected by users. Based on the gap model, SERVQUAL 1
c 1 1
ð1rÞMmt
score (represented by SQS) is utilized to model the CPV for ¼ tMmPM e dt   r fW ðtÞ dt
0 s0 s 0
a service, and is expressed as !
PM PM c 1
X
n ¼   ð1  Pb Þ  r:
SQS ¼ ws ðPecs  Exps Þ; (16) ð1  rÞ2 Mm 1  r s0 s
s¼1
(19)

SQSprice ðrÞ ¼ SQSprice ðr; W Þ ¼ Pecprice  Expprice SQSwaitingtime ¼ SQSwaitingtime ðrÞ

  Z 1
Z c 1 1 z
¼ Cðr; W Þ  ar ¼
s0 s
ðar  arÞ dt
¼ e r SQSwaitingtime ðzÞ dz
r
  0
0
!
Z a þ c 1 ! Z 1
 PM PM c 1
þ  
g s0 s
fW ðtÞ
cg g
 r  gt dt þ   arfW ðtÞ dt ¼   ð1  Pb Þ  r:
c 1 s0 s a c 1 ð1  rÞ2 Mm 1  r s0 s
s0  s g þs 0  s
  !
Z a þ c 1 (20)
g s0 s cg g
¼    rMmPM eð1rÞMmt dt
c 1
s0  s
s0 s where n is the number of the service’s attributes, ws is the
 
Z a þ c 1 weight of the service’s sth attribute, Pecs is the perception
g s0 s
   MmPM gteð1rÞMmt dt for the attribute, and Exps is the expectation for the attribute.
c 1
s0  s In this paper, we consider two most important attributes
Z 1
(i.e., price and waiting-time) of a service and treat their impor-
   arMmPM eð1rÞMmt dt
a c 1
g þs 0  s
tance as equal. In such a case, we have n ¼ 2, w1 ¼ w2 ¼ 0:5,
! ! Pec1 ¼ Pecprice , Exp1 ¼ Expprice , Pec2 ¼ Pecwaitingtime , and
PM r cg g a c 1 c 1
¼  eð1rÞMmðg þs0 sÞr  eð1rÞMmðs0 sÞr Exp2 ¼ Expwaitingtime . For the price attribute, the perception
r1 s0 s
! ! Pecprice is the service charge Cðr; W Þ given in Eq. (8) and the
 
PM g a c 1 a c 1 expectation Expprice is ar while for the waiting time attribute,
  ð1 þ ð1  rÞMm þ  r eð1rÞMm g þs0 s r
r1 g s0 s the perception Pecwaitingtime is W and the expectation
! !  ! Expwaitingtime is ðsc0  1sÞr. Using the gap model, we can derive
c 1 ð1rÞMm sc 1s r
 1 þ ð1  rÞMm  r e 0 the SQS with respect to price and waiting time attributes (rep-
s0 s
  resented by SQSprice and SQSwaitingtime ) as in Eqs. (3.1)-(14).
arPM ð1rÞMm gaþsc 1s r (See the top of this page.) Note that the derived SQSprice and
þ e 0 :
r1 SQSwaitingtime are with different units. To consider the two
(17) attributes together, we need to normalize the value of the two
0 0
variables. Let SQSprice and SQSwaitingtime represent the nor-
Z 1 malized SQSprice and SQSwaitingtime obtained by the maxi-
1 z
SQSprice ¼ SQSprice ðrÞ ¼ e r SQSprice ðzÞdz mum-minimum normalization method [34], respectively.
0 r
Then, the SERVQUAL score SQS can be calculated as
PM 1
¼ 0 0
ðr  1Þr ðr þ ð1  rÞMmðga þ sc0  1sÞÞ2 SQS ¼ 0:5  SQSprice þ 0:5  SQSwaitingtime : (21)
! !
1 cg g
   We adopt a user demand function [35] which indicates that
ðr þ ð1  rÞMmðsc0  1sÞÞ2 s0 s
the user’s demand for services is linear with the CPV (quan-
!
PM g a c 1 tified by SQS). That is
 r þ ð1  rÞMm þ 
ðr  1Þr g s0 s DCPV ðSQSÞ ¼ c þ f  SQS; (22)
!
1 c 1
  2  ðr þ ð1  rÞMm  where c and f ðc; f > 0Þ are constants representing the
s0 s
þ ð1  rÞMm g þ s0  s
1
r
a c 1
basic demand and potential market demand. Note that the
!! calculation of SQS is independent of the user demand func-
1 aPM
 1 þ tion used here.
1 2
ðr þ ð1  rÞMmðs  sÞÞ
c ðr  1Þr
0
1 3.2 Validate SERVQUAL Model and Demand
  2 : Function
r þ ð1  rÞMm g þ s0  sÞ
1 a c 1
To validate the SERVQUAL model and user demand func-
(18) tion, we carry out several simulations to evaluate SQSprice ,
WANG ET AL.: CUSTOMER PERCEIVED VALUE- AND RISK-AWARE MULTISERVER CONFIGURATION FOR PROFIT MAXIMIZATION 1079

defines a function Y ¼ c þ f  SQS   and finds a mono-

tone interval ½low ; high  such that Y ðlow Þ > 0 and
Y ðhigh Þ < 0 hold (line 1). It then iteratively divides the inter-
val in two by computing the mid point mid ¼ ðlow þ high Þ=2
of the interval and the value of function Y at that point if
Y ðlow Þ  Y ðhigh Þ > " holds for an arbitrarily small positive
number " (lines 2-9). In each round of iteration, the searching
interval ½low ; high  and the mid point mid are updated.
When the iteration stops, the algorithm selects the updated
mid point as the DCPV ðSQSÞ (lines 10-11).

Algorithm 1. The SB-Based Method to Derive User

Demand
1: define a function Y ¼ c þ f  SQS   and find a
monotone interval ½low ; high  such that Y ðlow Þ > 0
and Y ðhigh Þ < 0;
2: while Y ðlow Þ  Y ðhigh Þ > " do
3: mid ¼ ðlow þ high Þ=2;
4: if Y ðmid Þ < 0 then
5: high ¼ mid ;
6: else
7: low ¼ mid ;
8: break;
9: calculate Y ðlow Þ and Y ðhigh Þ;
10: mid ¼ ðlow þ high Þ=2;
11: DCPV ðSQSÞ ¼ mid ;
Fig. 1. SQSprice , SQSwaiting-time , and DCPV for different service arrival
rates ().
4 PROFIT OPTIMIZATION SCHEME
SQSwaitingtime , and DCPV for different service arrival rates In this section, we describe the proposed profit optimization
(). In the simulations, c is set to 15 and f is set to 20 as in scheme in detail. Specifically, we first show the method to
[35]. The normalized results of SQSwaitingtime and SQSprice calculate the cloud service provider’s profit, then discuss the
are summarized in Fig. 1a. From the figure, we can clearly optimal multiserver configuration for maximizing the profit,
observe that with the increase in the number of arrival serv- and finally present the heuristic algorithm to obtain the max-
ices, SQSwaitingtime and SQSprice both remain the same when imum profit considering the risk in the pricing contract.
the service number is less than 11 and then
SQSwaitingtime =SQSprice significantly increases/decreases 4.1 Calculate the Cloud Service Provider’s Profit
when the service number is larger than 11. This observation The profit of a cloud service provider is generated from the
is in line with the SLA model that when users’ waiting time gap between the revenue earned by purchasing services to
exceeds a threshold, the service provider will reduce the ser- customers and the monetary cost of processing customers’
vice charge/price to compensate for the latency. The results service requests. Below, we describe the methods to esti-
of DCPV are summarized in Fig. 1b. It can be easily observed mate the revenue, cost, and profit in detail.
from the figure that with the increase in the number of arrival Revenue. To estimate the revenue R of a cloud service
services, DCPV remains the same when the service number is provider, we first need to know the expected charge Cexp to
less than 11. When the service number is larger than 11, a service request. The Cexp is calculated as follows [4]:
DCPV increases first and then decreases. This observation is Z 1
in line with the phenomenon in the cloud service market that CðrÞ ¼ Cðr; W Þ ¼ fW ðtÞCðr; tÞ dt (23)
1
when users’ waiting time exceeds a threshold, the user ser- Z 1
vice demand will increase first due to the compensation and Cexp ¼ CðrÞ ¼ fr ðzÞCðzÞdz; (24)
then decrease due to the intolerable waiting time. The two 0
observations discussed above show the correctness of the where fW ðtÞ is the PDF of waiting time W given in Eq. (5),
SERVQUAL model and user demand function. fr ðzÞ is the impulse function given in Eq. (6), and
Cðr; W Þ=Cðr; tÞ is given in Eq. (8). Using Eq. (2.3) and the opti-
3.3 Derive User Demand Based on CPV
mal setting of parameter g (i.e., g ¼ a=ðc=s0  1=sÞ [4]), Cexp is
Using Eqs. (21)-(22) to compute DCPV ðSQSÞ directly is chal- derived as
lenging due to the complexity of SQSprice and SQSwaitingtime .
However, by using the standard bisection (SB) method, we Pb ar
can easily find a numerical solution of DCPV ðSQSÞ. There- Cexp ¼ 
ð2ðMs  rÞðsc0  1sÞ þ 1ÞððMs  rÞðsc0  1sÞ þ 1Þ
fore, we develop a SB-based algorithm to obtain the user
demand DCPV ðSQSÞ. The details of our algorithm are given þ ar ¼ CðM; sÞ:
in Algorithm 1. The algorithm works as follows. It first (25)
1080 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 31, NO. 5, MAY 2020

Clearly, Cexp can be described as a function of the number 4.2 Derive the Optimal Setup to Maximize the Profit
and speed of servers (i.e., M and s). As introduced in Section 2.2, PM is the probability of having
The total revenue of a cloud service provider is the prod- exact M service requests in the system and can be derived
uct of the expected charge to a service request and the user using Eq. (2) as
demand for this service, i.e.,
ðMrÞM
Revenue ¼ CðM; sÞ  DCPV ; (26) PM ¼ P0 :
M!
P
By applying Taylor series expansions M1 k
k¼0 ðMrÞ =k!  e
Mr
where DCPV is the user demand for this service based on pffiffiffiffiffiffiffiffiffiffiffi M M
customer perceived value. and M!  2pM ð e Þ , PM is then approximately expressed
Cost. The cost of a cloud service provider contains two as
parts: the money paid to rent cloud computing infrastructure 1r
and the electricity expense to maintain the operation of the PM  pffiffiffiffiffiffiffiffiffiffiffi r1 :
2pM ð1  rÞðe r ÞM þ 1
computing infrastructure. The cost model is a commonly
accepted model that has been widely used in the litera-
Substituting the approximate PM and the Pb given in Eq. (4)
ture [1], [4], [5], [6], [20], [26]. Suppose d is the fee paid by the into CðM; sÞ, we can derive
provider to rent a server during a sale period T . For renting a
 1
multiserver system composed of M servers during the CðM; sÞ  ar 1   
period T , the provider then needs to pay 2ðMs  rÞðsc0  1sÞ þ 1
1 1 
Rent ¼ Md  T: (27)     pffiffiffiffiffiffiffiffiffiffi
ffi  :
ðMs  rÞðsc  1sÞ þ 1 r
2pM ð1  rÞðeerÞM þ 1
0
Electricity fee is a significant expense for today’s data cen-
ters and can be calculated as the product of servers’ energy For the sake of easy representation, we let
consumption and electricity price. The energy consumption pffiffiffiffiffiffiffiffiffiffiffi
f1 ¼ 2pM ð1  rÞðer =erÞM þ 1;
of a server can be modeled at different levels of abstraction.

In this paper, we consider server’s energy consumption at c 1
the abstraction level of digital CMOS circuit. Let E T repre- f2 ¼ 2ðMs  rÞ  þ 1;
s0 s
sent the energy consumption of the M servers during the 
sale period T . Then, using the CMOS-level power model c 1
f3 ¼ ðMs  rÞ  þ 1:
introduced in [5], E T can be calculated as s0 s

E T ¼ M  ðPowdyn  r þ Powsta Þ  T; (28) CðM; sÞ is then rewritten as


1
where r is the server utilization, Powdyn is the dynamic CðM; sÞ ¼ ar 1  :
f1 þ f2 þ f3
power dissipation, and Powsta is the static power dissipa-
tion. Assuming C T is the price of the energy consumed by
From Eq. (1) we can deduce that Mr ¼ DCPV r
s . Substitut-
servers during the period T , the electricity bill is derived as
ing it into CðM; sÞ and GðM; sÞ, then we can find the optimal
M for maximizing profit by calculating the partial deriva-
Bill ¼ E T  C T ¼ M  ðPowdyn  r þ Powsta Þ  T  C T :
tive of GðM; sÞ with respect to M. That is
(29)
@G @C
Increasing server utilization is helpful to serve more requests ¼ DCPV  ðd þ Psta  C T Þ  T:
@M @M
and hence bring more revenue, but it also leads to a higher
@G
electricity bill. The service provider obtains more profit only The partial derivatives used for calculating @M are
when the extra revenue is greater than the additional electric- !
ity bill. @C ar @f1 @f2 @f3
¼  f2  f3 þ  f1  f3 þ  f1  f2 ;
Profit. The cloud service provider’s profit during a sale @M ðf1  f2  f3 Þ2 @M @M @M
period T is derived as the revenue minus the expenses
!
including the rental cost and electricity cost, i.e., @f1 pffiffiffiffiffiffi 1 r pffiffiffiffiffi
¼ 2p pffiffiffiffiffi ð1  rÞC þ pffiffiffiffiffi C þ M ð1  rÞ C ;
2

Profit ¼ Revenue  Rent  Bill; (30) @M 2 M M

! !
where Revenue, Rent, and Bill are given in Eqs. (23), (24), @f2 c 1 @f3 c 1
¼ 2s  ; ¼s  ; C ¼ ðer =erÞM :
and (26), respectively. Substituting these into Eq. (27), we @M s0 s @M s0 s
can rewrite Profit as a function, represented by GðM; sÞ, of
the multiserver configuration M and s. That is Similarly, we can find the optimal s for maximizing profit
by calculating the partial derivative of GðM; sÞ with respect
Profit ¼ CðM; sÞ  DCPV  Md  T to s. That is
 MðPowdyn  r þ Powsta Þ  C T  T ¼ GðM; sÞ: @G @C
¼ DCPV  DCPV rða  1Þsa2  C T  T:
(31) @s @s
WANG ET AL.: CUSTOMER PERCEIVED VALUE- AND RISK-AWARE MULTISERVER CONFIGURATION FOR PROFIT MAXIMIZATION 1081

The partial derivatives used for calculating @G

@s are decisions at the beginning of each scheduling interval. The
! problem is formulated as
@C ar @f1 @f2 @f3
¼ f2 f3 þ f1 f3 þ f1 f2 ; Max : Profitrisk ¼ GðM; sÞ þ RISKi : (35)
@s ðf1 f2 f3 Þ2 @s @s @s i

@f1 pffiffiffiffiffiffiffiffiffiffiffi  @C @C M From Eqs. (28)-(31) we can see that the key of solving the
¼ 2pM r þ Mð1  rÞ2 ; ¼ ð1  rÞC;
@s @s @s s maximization problem is to find the optimal point of
! GðM; sÞ. Solving the optimization function GðM; sÞ analyti-
@f2 Mc DCPV r @f3 Mc DCPV r cally requires that M and s are continuous variables and
¼2  ; ¼  :
@s s0 s2 @s s0 s2 needs to introduce a closed-form expression for approximat-
ing the function [4], [5], [6]. However, considering that M
Once the partial derivatives of GðM; sÞ given in Eq. (4.1) and s of actual systems are discrete variables and the error
with regard to M and s are obtained, we can compute and caused by closed-form approximation is non-negligible, ana-
derive the optimal solutions to maximizing profit by setting lytically solving the function is not feasible. Although we
these partial derivatives equal to 0. cannot solve the function analytically, we still can carry out
extensive experiments to plot out GðM; sÞ under varying
4.3 Risk-Aware Profit Maximization configurations for checking the existence of optimal solution
Although finding an optimal configuration of M and s is help- to GðM; sÞ as in [4], [6]. Fig. 2 presents the values of GðM; sÞ
ful to maximize the profit, the pricing contract still involves under six different configurations. All the figures of GðM; sÞ
some risk for the cloud service provider. Specifically, due to show that there must be an optimal point where the profit is
the inherent uncertainty in the probability distribution of maximized. Note that the experiment results here are not
waiting times and the number of task instructions, the esti- meant to be comprehensive but just to demonstrate that the
mated distributions of time and cost may be different from optimal point exists. Based on this observation, we then use
the actual ones, leading to a decreased profit or even a loss. To a heuristic algorithm to find a numerical optimal solution.
make the studied problem more realistic, we formally define
the risk based on loss and take into account the risk by adding Algorithm 2. Calculate the Risk-Aware Profit
it as a part of the optimization objective. The definitions of 1: if i ¼ 1 then
LOSS, RISK, and Risk-aware Profit are described as below. 2: LOSSi ¼ 0;
3: RISKi ¼ 0;

Definition 1: LOSS. For the ith running of a specific 4: calculate Profitrisk using Eq. (30);
service, let Mi and si represent the actual number of 5: else
i

servers and the actual operating speed of servers, 6: j ¼ 1;

respectively. When the cloud service provider gener- 7: LOSStotal ¼ 0;
ates a contract with his/her predicted configuration 8: while j 4 i  1 do
M and s, the loss of profit is then derived as 9: LOSSj ¼ GðMj ; sj Þ  GðM; sÞ;
10: LOSStotal ¼ LOSStotal þ LOSSj ;
LOSSi ¼ GðMi ; si Þ  GðM; sÞ; (32) 11: j þ þ;
12: calculate RISKj using Eq. (29);
where G is the function to calculate the profit.
13: calculate Profitrisk using Eq. (30);

Definition 2: RISK. For the ith running of a specific i

service, the risk of a service provider is defined as

To find the numerical optimal solution, we first need to
the average of profit loss over the previous i  1 runs
discretize the solution space. Under the discretized solution
of this service. That is
space, the simplest way to find the optimal solution is using
Pi1 brute force search, which is however not suitable because of
j¼1LOSSj
RISKi ¼ : (33) its high time complexity. To overcome this shortcoming, we
i1 propose a discrete simulated annealing (SA)-based heuristic

Definition 3: Risk-aware Profit. Considering the risk, algorithm. The SA [36] is motivated by an analogy to physi-
the profit gained by providing a service is derived as cal annealing in solids. It is a probabilistic technique for
approximating the global optimum of a given function. It
Profitrisk ¼ GðM; sÞ þ RISKi : (34) can avoid being trapped at local optima by introducing a
i
diversification mechanism (also called restart strategy). The
We summarize the basic steps needed to compute the risk- details of our heuristic algorithm are given in Algorithm 3.
aware profit in Algorithm 2. After introducing the risk-related The algorithm begins with an initial solution obtained by
concepts and how to estimate the risk-aware profit, we then assuming the multiserver configuration with the maximum
present the risk-aware profit maximization problem, followed capacity is utilized (lines 1-5). Note that the effectiveness of
by our approach to the profit maximization problem. the algorithm is independent of the initial solution selection.
In this paper, we aim at maximizing the profit of the Any arbitrary valid initial solution can be also used here. It
cloud service provider with the consideration of the risk in then initializes the “temperature” of this solution as a high
pricing contract by determining the optimal number of serv- value, sets the Boltzmann constant k, the cooling rate CR, the
ers, operating speed of servers, and price of services. We iteration number iteratornum , and the changing steps (stepM
assume that the cloud service provider optimizes his/her and steps ) of multiserver size and speed, and generates a
1082 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 31, NO. 5, MAY 2020

Fig. 2. The values of GðM; sÞ under varying configurations.

random constant $thresh from the range of (0,1] (lines 6-10). (i.e., jProfitnext  Profitcur j < " where " is an arbitrarily small
The “temperature” will gradually decrease during the simu- positive number), the outer while-loop is terminated and the
lated annealing process. At each temperature, the algorithm global optimal solution is found (lines 35-36).
performs a certain amount of iterations and considers the Fig. 3 shows the searching process of the optimal solution
neighbor solutions. Once a new solution is obtained (lines 13- using Algorithm 3. The blue points in the figure are the local
18), its profit (represented by Profitnext ) is calculated using optimal solutions obtained during the search and the red
Eq. (4.1) and compared to that (represented by Profitcur ) of point (M ¼ 20:78; s ¼ 2:995; Profit ¼ 202:4) is the final solu-
the current solution (lines 19-20). If DProfit ¼ Profitnext  tion found by Algorithm 3. From the figure we can deduce
Profitopt > 0, indicating the new solution can bring a higher that the algorithm can not only avoid falling into local optima
profit, the new solution is accepted (lines 21-25). Otherwise, but also find a solution that is very close to the global optimal
the new solution is a worse one and the algorithm then choo- solution.
ses to either keep the current solution (lines 31-33) or accept
the new one with a probability (lines 26-30), in order to avoid
getting stuck in local optima. During the simulated annealing 5 NUMERICAL RESULTS
process (i.e., the optimal solution searching process), the In this paper, we consider the scenario that a cloud service
“temperature” is reduced using a cooling coefficient (line 34). provider rents a multiserver system to serve the requests
After a large amount of iterations for searching the optimal submitted by cloud users. Such a multiserver system could
solution, if the profits of neighbor solutions become very close be in the multiple forms of infrastructures, e.g., blade centers
[23], clusters of traditional servers [24], and multicore server
processors [25]. The cloud service provider charges users for
processing their service requests based on the services
amount and quality. To represent this scenario, we need to
simulate a multiserver system and estimate the service
demand of users. In the experiment, we assume that the mul-
tiserver system is in the form of multicore server processors,
and thus we use several up-to-date multiprocessors (i.e.,
Intel Xeon Skylake-Platinum 8153 [37], Intel KNL-Xeon Phi
7250 [38], NVIDIA-V100 [39], IBM Power7-BladeCenter
PS704 [40]) to simulate the multiserver system. The parame-
ters used to estimate the electricity cost and renting cost of
this multiserver system as well as the parameters related to
service requests are listed in Table 1. The values of these
parameters (such as the multiserver size M, the multiserver
Fig. 3. An illustration of the process of searching the optimal solution speed s, and the user service demand DCPV ) adopted in the
using Algorithm 3. experiments are either extracted from actual system (i.e.,
WANG ET AL.: CUSTOMER PERCEIVED VALUE- AND RISK-AWARE MULTISERVER CONFIGURATION FOR PROFIT MAXIMIZATION 1083

TABLE 1
Definition of Main Notations Used in the Experiment
Notation Definition Value
s0 the customer’s expected service processing speed 1 billion instructions per second [4]
c a constant coefficient 3 [4]
r the expected number of instructions to complete a service 1 billion instructions [4]
a the service charge per unit amount of service 15 cents per billion instructions [4]
d the fee of renting a server during a sale period 1.5 cents per second [4]
 the processor dependent coefficient 9.4192 [4]
a a constant coefficient 2 [4]
Powsta the static power dissipation 2 Watts per second [5]
CT the price of the energy consumed by the servers during a sale period 0.1 cents per Watt  second [5]

Intel Xeon Skylake-Platinum 8153, Intel KNL-Xeon Phi 7250, Algorithm 3. The Discrete SA-Based Heuristic to
NVIDIA-V100, IBM Power7-BladeCenter PS704) or widely Maximize the Profit
accepted in the literature (e.g., [4], [5]). The values of M and s Input: DCPV ðSQSÞ, r, ½Mmin ; Mmax , ½smin ; smax ;
used in the experiments also satisfy the constraint of server Output: optimal server size Mopt , optimal server speed sopt ,
utilization, i.e., 0 < r ¼ Mm 
¼ Mðs=rÞ

¼ Ms
r
< 1. Thus, our and maximal profit Profitopt ;
parameter settings should be realistic and acceptable. 1: discretize ½Mmin ; Mmax  and ½smin ; smax ;
In this section, three sets of simulation experiments are 2: select node ðMmax ; smax Þ as start node ðM; sÞ;
conducted to validate the proposed scheme from different 3: Mcur ¼ Mmax , scur ¼ smax ;
perspectives. In the first set of experiments, we observe the // the initial point;
changing trends of user demand and profit under varying 4: calculate Profitcur ¼ GðMcur ; scur Þ using Eq. (4.1);
multiserver configurations. This experiment is to show how 5: initialize Mopt ¼ 0, sopt ¼ 0, Profitopt ¼ 0;
the user demand and profit are affected by multiserver con- 6: set temperature ¼ 2000;
figurations. In the second set of experiments, we find the opti- // the initial temperature;
mal configuration of multiserver speed and size as well as the 7: set k ¼ Boltzmann constant;
corresponding profit under different user demands [41]. This 8: set CR ¼ 0:99; // the cooling rate;
experiment is to show how the optimal multiserver configu- 9: set iteratornum ¼ 10000, stepM ¼ 1, steps ¼ 0:1;
ration and profit are affected by user demands. In the third 10: generate a random constant from the range of (0,1] using
set of experiments, we compare the proposed profit maximi- $thresh ¼ Randð0; 1Þ;
zation scheme with two benchmarking methods. This experi- 11: while true do
ment is to show the effectiveness of the proposed scheme in 12: for i ¼ 0; i < iteratornum ; i þ þ do
13: while true do
increasing profit. All the simulation experiments are imple-
14: derive two variables by $M ¼ RandðintÞ and
mented on a laptop equipped with 2.2 GHz Intel i7 six-core
$s ¼ RandðdoubleÞ;
processor and 16 GB DDR4 memory, and running a Unix ver-
15: Mnext ¼ Mcur þ stepM  Mcur  $M ;
sion of Matlab x64.
16: snext ¼ scur þ steps  scur  $s ;
17: ifðMnext 2 ½Mmin ; Mmax && snext 2 ½smin ; smax Þ then
5.1 User Demand and Profit 18: Break;
In this set of experiments, we conduct a series of numerical 19: derive Profitnext ¼ GðMnext ; snext Þ by Eq. (4.1);
calculations to show the user demand and profit under the 20: DProfit ¼ Profitnext  Profitcur ;
varying configurations of the server size M and the server 21: if DProfit > 0 then
22: Mopt ¼ Mnext ; sopt ¼ snext ;
speed s, which take the value from f10; 15; 20; 25; 30; 35; 40;
23: Profitopt ¼ Profitnext ;
45; 50; 55; 60; 65; 70g and f1; 1:25; 1:5; 1:75; 2:0; 2:25g, respec-
//keep the new solution;
tively. To derive the user demand and profit, we use the
24: Mcur ¼ Mnext ; scur ¼ snext ;
parameters s0 , r, a, and c as presented in Table 1. //the new point;
Fig. 4a shows that no matter which processing speed s is 25: Profitcur ¼ Profitnext ;
adopted by the multiserver system, the user demand for 26: DProfit
else if expðtemperaturek Þ > $thresh then
services first quickly increases to a peak, then gradually
27: Mcur ¼ Mnext ; scur ¼ snext ;
decreases, and finally keeps constant, as the multiserver size 28: Profitcur ¼ Profitnext ;
M grows from 10 to 70. The reason why the user demand //accept new point with a probability;
quickly increases first is that adding more servers can attract 29: Mopt ¼ Mnext ; sopt ¼ snext ;
more users by providing a higher computing performance. 30: Profitopt ¼ Profitnext ;
However, when the user demand becomes high, the multi- 31: else
server system may not be able to timely process these service 32: Mopt ¼ Mcur ; sopt ¼ scur ;
requests. In this case, users have to wait and their CPVs 33: Profitopt ¼ Profitcur ;
become low, which in turn decreases the demand of users //keep the current solution;
for service. Note that the user demand would eventually 34: temperature ¼ temperature  CR;
remain a constant. This is because there is always a certain 35: if jProfitnext  Profitcur j < " then
service demand in the market that is stable and not affected 36: Break;
by external factors.
1084 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 31, NO. 5, MAY 2020

Fig. 4. User demand and profit.

Fig. 5. Optimal speed and profit under varying multiserver sizes and user demands.

Similarly, we can observe from Fig. 4b that no matter which and the corresponding maximum profit for varying user
processing speed s is adopted by the multiserver system, the demands and multiserver sizes, then find the optimal multi-
cloud service provider’s profit also first rapidly increases to a server size and the corresponding maximum profit for vary-
peak, then gradually decreases, and finally keeps stable, as ing user demands and server speeds, and finally find the
the multiserver size M grows from 10 to 70. The initial optimal multiserver size and server speed under the given
increase of profit is due to the extra revenue gained by the user demands.
increased demand, which is much higher than the additional Fig. 5 shows the optimal server speed and the corre-
cost of using the newly added severs. However, with the con- sponding maximum profit when varying the multiserver
tinuous increase of server size, the extra revenue may not be size and user service demand. The server size takes the value
able to afford the additional cost and the user demand will from f10; 20; 30; 40; 50; 60; 70; 80g2 and the user service
decrease, leading to a reduced profit. Finally, the profit tends demand takes the value from f34:73; 30:23; 25:17; 19:7g. The
to be stable when the user demand becomes a constant. values of parameters s0 , r, a, c, z, a, Psta , d, C T are consistent
From Figs. 4a and 4b, we can also find that when the multi- with those in Table 1. As can be seen from Fig. 5a, for a given
server size is small (i.e., 10 M 20), increasing service service demand, the optimal server speed decreases with the
processing speed is effective in attracting more user demands increase in the size of the multiserver system. This is because
and hence obtaining a higher profit. But when the multiserver when the required computing capability for completing a
size becomes larger (i.e., M > 20), increasing service process- given amount of service requests is fixed, increasing the
ing speed cannot ensure more user demands and larger prof-
its. This is because that high server speed and large server size
both lead to a high cost of operating the multiserver system. 2. This setup could cover most of parameters of the aforementioned
up-to-date multiprocessors. However, it doesnot include the case of NVI-
5.2 Optimal Multiserver Configuration and Profit DIA-V100 (i.e., 5,120 cores) since such a large server size is not common in
real-world cloud service providers. For example, the server size used by
In this set of experiments, we carry out a series of numerical Amazon EC2 [2], Microsoft Azure [3], Aliyun [27], and Huawei Cloud [42]
computations in order to first find the optimal server speed are in the rang of [1, 96], [1, 64], [1, 104], and [1, 64], respectively.
WANG ET AL.: CUSTOMER PERCEIVED VALUE- AND RISK-AWARE MULTISERVER CONFIGURATION FOR PROFIT MAXIMIZATION 1085

Fig. 6. Optimal multiserver size and profit under varying server speeds and user demands.

TABLE 2
The Maximum Profit and the Corresponding Multiserver Configuration

Demand (DCPV ) 9 11 13 15 17 19 21 23 25 27
Optimal size 12 13 14 15 16 17 18 19 20 21
Optimal speed 1 1.1 1.2 1.3 1.35 1.38 1.43 1.44 1.45 1.46
Maximum profit 106.20 135.60 156.56 181.17 206.22 231.45 255.97 281.56 306.92 332.25

multiserver size necessitates the decreased server speed. demands. The user demand varies from 9 to 27 in the step of
Fig. 5b presents the corresponding maximum profit when 2. As can be seen from the table, the optimal multiserver size
increasing the multiserver size and user service demand. M is monotonically increasing with the increase of user
Clearly, larger user demands can bring higher profit for demands DCPV . The reason is that higher user demand needs
cloud service providers. From the figure we can also observe a greater computing capacity.
that the profit is affected by the multiserver size under a cer-
tain user demand. Specifically, when the multiserver size is 5.3 Performance Comparison
small and the server speed is high, renting more servers is In this set of experiments, we compare our CPV-based and
able to increase the profit. When the multiserver size risk-aware profit maximization scheme with two state of the
becomes greater further and exceeds a threshold, the profit art benchmarking methods COMCPM [4] and OMCPM [5]
decreases. This is because that to maintain a steady comput- under the same experimental settings. The two benchmark-
ing capability for the given user demand, the server speed ing methods both consider the service-level agreement as
necessarily decreases with the increase of multiserver size. well as customer satisfaction, and derive an optimal multi-
In the case of small multiserver size and high server speed, server configuration and service price for maximizing profit.
the saved energy cost derived by lowering the server speed Compared with COMCPM and OMCPM, our scheme is
is greater than the extra cost of renting more servers. How- based on the CPV that reflects users willingness to purchase
ever, when the multiserver size becomes large and the server cloud services and takes into account the risk involving in
speed drops to a certain level, the increased renting cost sur- the pricing contract. In the experiments, we vary the user
passes the saved energy cost, leading to a decreased profit. demands and find the optimal multiserver configuration (M
Fig. 6 presents the optimal multiserver size and the corre- and s) and the corresponding profit for each user demand.
sponding maximum profit when varying the server speed We use the same values of parameters such as s0 , r, a, c, z, a,
and user service demand. The server speed takes the value Psta , d, C T as in Table 1. The two benchmarking methods
from f1; 1:25; 1:5; 1:75; 2; 2:25; 2:5; 2:75; 3:0; 3:25; 3:5g and the COMCPM and OMCPM are briefly described below.
user service demand takes the value from f19:7; 25:17; 30:23;
34:73g. To derive the results, we use the parameters as same
COMCOM [4] is a profit maximization scheme that
in Table 1. From Figs. 6a and 6b, we can observe that with the specially considers the customer satisfaction in
increase of server speed, i) the optimal multiserver size cloud. The scheme operates as follows. It first defines
becomes smaller, which is due to that increasing the server the concept of customer satisfaction leveraged from
speed necessitates the smaller multiserver size under a cer- economics and develops a formula to measure the
tain user demand, and ii) the profit gradually decreases, customer satisfaction in cloud. Based on the affection
which is because that the high energy cost of using fast server of customer satisfaction on server’s workload, it then
speed leads to the decrease of profit. analyzes the interaction between the market demand
Table 2 presents the optimal multiserver size, optimal and the customer satisfaction as well as calculates
server speed, and maximum profit under varying user the service arrival rate. Finally, it solves the profit
1086 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 31, NO. 5, MAY 2020

TABLE 3
Profit Comparison Between COMCPM and the Proposed Scheme

COMCPM Proposed Improvement

ProposedCOMCPM
Demand optM optS Profit Demand optM optS Profit COMCPM
4.85 6 1.13 42.33 4.97 6 1.1 57.10 34.9%
9.79 12 1.04 88.68 9.96 10 1.19 118.18 33.3%
14.72 17 1.05 135.57 14.97 15 1.15 178.55 31.7%
19.68 22 1.06 182.71 19.96 20 1.12 239.46 31.1%
24.63 27 1.06 230.06 24.95 25 1.11 300.25 30.5%
29.59 32 1.06 277.55 29.97 30 1.09 361.21 30.1%
34.55 37 1.06 325.14 34.97 35 1.09 422.28 29.9%

TABLE 4
Profit Comparison Between OMCPM and the Proposed Scheme

OMCPM Proposed Improvement

ProposedOMCPM
Demand optM optS Profit Demand optM optS Profit OMCPM
4.78 6.05 1.05 42.04 4.94 6 1.08 56.18 33.7%
9.63 11.67 1.01 87.65 9.95 10 1.17 116.10 32.5%
14.55 17.23 0.99 134.75 14.93 15 1.14 176.37 30.9%
19.44 22.76 0.98 181.44 19.95 20 1.11 236.29 30.2%
24.43 28.27 0.98 229.49 24.92 25 1.10 297.71 29.7%
29.30 33.78 0.97 276.38 29.95 30 1.08 357.50 29.4%
34.31 39.27 0.97 324.65 34.95 35 1.08 419.07 29.1%

maximization problem for cloud service providers helpful for the cloud service provider to gain a higher
by finding out the optimal configuration of servers. profit.
In COMCOM, the optimal configuration of servers is
obtained using a discrete hill climbing algorithm.

OMCPM [5] is a profit maximization scheme that 6 CONCLUSION
treats a multiserver system as an M/M/M queuing In this paper, we consider customer perceptive value and
model as well as considers two server speed and risk in solving the optimal multiserver configuration prob-
power consumption models. The scheme works as lem for maximizing the cloud service provider’s profit. Spe-
follows. It first derives the probability density func- cifically, we propose a method to model the customer
tion (PDF) of the waiting time of a newly arrived ser- perceptive value and hence estimate the user demand for
vice request and calculates the expected service cloud services. Based on the estimated user demand, we
charge to a service request. Based on the PDF and study the optimal multiserver configuration problem for
expected service charge of service requests, it then profit maximization and show how to derive the analytical
gets the expected net business gain (i.e., profit) in optimal solution. Furthermore, we develop a simulated
one unit of time and obtain the optimal configuration annealing based algorithm to find the numerical optimal
of server size and server speed numerically. solution when considering the risk in the pricing contract. To
Table 3 compares the profit achieved by the proposed validate the effectiveness of the proposed scheme, we ana-
scheme with that of COMCPM [4]. It is clear that the pro- lyze the changing trend of profit derived by the proposed
posed scheme outperforms COMCPM in terms of increas- scheme under different configurations and compare the pro-
ing the cloud service provider’s profit. The improvement posed scheme with two benchmarking approaches. The
of profit achieved by the proposed scheme over COMCPM results reveal that the proposed scheme not only follows the
is 31.6 percent on average, and can be up to 34.9 percent. supply and demand law of the cloud service market but also
We can also derive a similar conclusion from Table 4 that increases the profit by up to 34.9 percent when compared to
the profit of the proposed scheme is greater than that of the two benchmarking approaches.
OMCPM [5]. Specifically, the profit improvement achieved
by the proposed scheme over OMCPM is 30.8 percent on
average, and can be up to 33.7 percent. The higher profit ACKNOWLEDGMENTS
achieved by the proposed scheme benefits from not only The authors would like to thank the editors and the anony-
the optimal configuration of servers (i.e., the server size M mous reviewers for their constructive comments and valuable
and the server speed s) but also the more user service suggestions. In addition, the authors would like to express
demands captured by the scheme. In addition, the pro- their sincere gratitude to their collaborator, professor Keqin
posed scheme also considers the risk involving in the pric- Li in the State University of New York at New Paltz, for his
ing contract to avoid the loss caused by the risk. This is helpful advice on this work. This work was supported in part
WANG ET AL.: CUSTOMER PERCEIVED VALUE- AND RISK-AWARE MULTISERVER CONFIGURATION FOR PROFIT MAXIMIZATION 1087

by the National Natural Science Foundation of China (Grant [21] Google Compute Engine, 2019. [Online]. Available: https://cloud.
google.com/products/compute-engine
Nos. 61802185 and 61272420), in part by Natural Science Foun- [22] Amazon CloudFormation, 2019. [Online]. Available: http://aws.
dation of Jiangsu Province (Grant No. BK20180470), in part by amazon.com/cloudformation
the Fundamental Research Funds for the Central Universities [23] K. Li, “Optimal load distribution for multiple heterogeneous
(Grant No. 30919011233), in part by Open Research Project of blade servers in a cloud computing environment,” J. Grid Comput.,
vol. 11, no. 1, pp. 27–46, 2013.
The Hubei Key Laboratory of Intelligent Geo-Information [24] B. N. Chun and D. E. Culler, “User-centric performance analysis
Processing (Grant No. KLIGIP-2018A04), and in part by of market-based cluster batch schedulers,” in Proc. IEEE/ACM Int.
Shanghai Municipal Natural Science Foundation under Grant Symp. Cluster Comput. Grid, 2002, pp. 30–30.
[25] K. Li, “Optimal configuration of a multicore server processor for
16ZR1409000. managing the power and performance tradeoff,” J. Supercomput.,
vol. 61, no. 1, pp. 189–214, 2012.
[26] K. Li, J. Mei, and K. Li, “A fund-constrained investment scheme
REFERENCES for profit maximization in cloud computing,” IEEE Trans. Services
[1] P. Cong et al., “Developing user perceived value based pricing Comput., vol. 11, no. 6, pp. 893–907, Nov./Dec. 2018.
models for cloud markets,” IEEE Trans. Parallel Distrib. Syst., vol. 29, [27] Aliyun, 2019. [Online]. Available: https://www.aliyun.com/
no. 12, pp. 2742–2756, Dec. 2018. price/product?spm=5176.8789780.1092586.1.1da155caFpHojQ&
[2] Amazon EC2, 2019. [Online]. Available: https://aws.amazon. aly_as=6NLKdXpV#/ecs/detail
com/cn/ec2/pricing/on-demand [28] Example of queuing models, 2018. [Online]. Available: http://
[3] Microsoft Azure, 2019. [Online]. Available: https://azure. www.universalteacherpublications.com/univ/ebooks/or/Ch10/
microsoft.com/zh-cn/pricing/details/virtual-machine-scale-sets mm1ex.htm
[4] M. Jing, K. Li, and K. Li, “Customer-satisfaction-aware optimal [29] L. Kleinrock, Queueing Systems: Computer Applications, vol. 2. New
multiserver configuration for profit maximization in cloud York, NY, USA: Wiley, 1976.
computing,” IEEE Trans. Sustain. Comput., vol. 2, no. 1, pp. 17–29, [30] R. C. Lewis and B. H. Booms, “The marketing aspects of service
Jan.–Mar. 2017. quality,” Emerg. Perspectives Services Marketing, vol. 65, no. 4,
[5] J. Cao, K. Huang, K. Li, and A. Y. Zomaya, “Optimal multiserver pp. 99–107, 1983.
configuration for profit maximization in cloud computing,” IEEE [31] M. Unuvar, S. Tosi, Y. Doganata, M. Steinder, and A. Tantawi,
Trans. Parallel Distrib. Syst., vol. 24, no. 6, pp. 1087–1096, Jun. 2013. “Selecting optimum cloud availability zones by learning user satis-
[6] M. Jing, K. Li, A. Ouyang, and K. Li, “A profit maximization faction levels,” IEEE Trans. Service Comput., vol. 8, no. 2, pp. 199–211,
scheme with guaranteed quality of service in cloud computing,” Mar./Apr. 2015.
IEEE Trans. Comput., vol. 64, no. 11, pp. 3064–3078, Nov. 2015. [32] A. Parasuraman, V. A. Zeithaml, and L. L. Berry, “SERVQUAL: A
[7] Y. Lee, C. Wang, A. Y. Zomaya, and B. Zhou, “Profit-driven schedul- multiple-item scale for measuring consumer perc,” J. Retailing,
ing for cloud services with data access awareness,” J. Parallel Distrib. vol. 64, no. 1, 1988, Art. no. 12.
Comput., vol. 72, no. 4, pp. 591–602, 2012. [33] L. S. V. Velsen, M. F. Steehouder, and T. D. Jong, “Evaluation of user
[8] H. Chun, “Optimal pricing and ordering policies for perishable support: Factors that affect user satisfaction with helpdesks and
commodities,” Eur. J. Oper. Res., vol. 144, no. 1, pp. 68–82, helplines,” IEEE Prof. Commun., vol. 50, no. 3, pp. 219–231, Sep. 2007.
Jan. 2003. [34] Normalization, 2019. [Online]. Available:https://en.wikipedia.
[9] Amazon Web Services, 2019. [Online]. Available: http://aws. org/wiki/Normalization_(statistics)
amazon.com/ [35] Y. Wang, A. Meliou, and G. Miklau, “A consumer-centric market
[10] Google App Engine, 2019. [Online]. Available: https://appengine. for database computation in the cloud,” 2016. [Online]. Available:
google.com/ https://arxiv.org/abs/1609.02104
[11] M. Ghamkhari and H. Mohsenian-Rad, “Energy and performance [36] S. Lyden and M. E. Haque, “A simulated annealing global maxi-
management of green data centers: A profit maximization mum power point tracking approach for PV modules under par-
approach,” IEEE Trans. Smart Grid, vol. 4, no. 2, pp. 1017–1025, tial shading conditions,” IEEE Trans. Power Electron., vol, 31, no. 6,
Jun. 2013. pp. 4171–4181, Jun. 2016.
[12] Amazon EC2 spot instances, 2019. [Online]. Available: https:// [37] Intel Xeon Skylake, 2017. [Online]. Available: https://ark.intel.
aws.amazon.com/cn/ec2/spot/pricing com/content/www/us/en/ark/products/series/125191/intel-
[13] H. Xu and B. Li, “Dynamic cloud pricing for revenue maxi- xeon-scalable-processors.html
mization,” IEEE Trans. Cloud Comput., vol. 1, no. 2, pp. 158–171, [38] Intel KNL, 2019. [Online]. Available: https://en.wikipedia.org/
Jul.–Dec. 2013. wiki/Xeon_Phi
[14] M. Macias and J. Guitart, “A genetic model for pricing in cloud com- [39] NVIDIA V100, 2019. [Online]. Available: https://en.wikipedia.
puting markets,” ACM Symp. Appl. Comput., pp. 113–118, 2011. org/wiki/List_of_Nvidia_graphics_processing_units
[15] S. Ren and M. V. D. Schaar, “Dynamic scheduling and pricing in [40] IBM Power7, 2019. [Online]. Available: https://en.wikipedia.org/
wireless cloud computing,” IEEE Trans. Mobile Comput., vol. 13, wiki/POWER7
no. 10, pp. 2283–2292, Oct. 2014. [41] Y. Hu, J. Wong, G. Iszlai, and M. Litoiu, “Resource provisioning
[16] S. Chatterjee, R. Ladia, and S. Misra, “Dynamic optimal pricing for for cloud computing,” in Proc. Conf. Center Adv. Stud. Collabortive
heterogeneous service-oriented architecture of sensor-cloud infra- Res., 2009, pp. 101–111.
structure,” IEEE Trans. Services Comput., vol. 10, no. 2, pp. 203–216, [42] Huawei Cloud, 2019. [Online]. Available: https://www.
Mar./Apr. 2017. huaweicloud.com/product/ecs.html
[17] V. A. Zeithaml, “Consumer perceptions of price, quality, and
value: A means-end model and synthesis of evidence,” J. Market-
ing, vol. 52, no. 3, pp. 2–22, 1988. Tian Wang (S’19) is currently working toward the
[18] Z. Yang and R. T. Peterson, “Customer perceived value, satisfac- PhD degree in the School of Computer Science
tion, and loyalty: The role of switching costs,” Psychol. Marketing, and Engineering, Nanjing University of Science
vol. 21, no. 10, pp. 799–822, 2004. and Technology, Nanjing, China. His current
[19] H. G. M. Gordon and L. Terrence, “Customer satisfaction with research interests include the areas of cloud
services: Putting perceived value into the equation,” J. Services computing and cyber security. He is a student
Marketing, vol. 14, no. 5, pp. 392–410, 2000. member of the IEEE.
[20] P. Cong et al., “User perceived value-aware cloud pricing for
profit maximization of multiserver systems,” in Proc. IEEE Int.
Conf. Parallel Distrib. Syst., 2017, pp. 537–544.
1088 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 31, NO. 5, MAY 2020

Junlong Zhou (S’15–M’17) received the PhD Tongquan Wei (M’11–SM’19) received the PhD
degree in computer science from East China Nor- degree in electrical engineering from Michigan
mal University, Shanghai, China, in 2017. He was Technological University, Houghton, Michigan, in
a visiting scholar with the University of Notre 2009. He is currently an associate professor with
Dame, Notre Dame, Indiana, from 2014 to 2015. the Department of Computer Science and Technol-
He is currently an assistant professor with the ogy, East China Normal University, Shanghai,
School of Computer Science and Engineering, China. His current research interests include real-
Nanjing University of Science and Technology, time embedded systems, green and reliable com-
Nanjing, China. His research interests include puting, parallel and distributed systems, and cloud
real-time embedded systems, cloud computing computing. He has published more than 70 papers
and IoT, and cyber physical systems. He has pub- in these areas, most of which are published in pre-
lished 60 refereed papers, most of which are published in premium con- mium conferences and journals including IEEE/ACM DAC, IEEE/ACM
ferences and journals. He received the Reviewer Award from the Journal ICCAD, IEEE/ACM DATE, the IEEE Transactions on Parallel and Distrib-
of Circuits, Systems, and Computers, in 2016. He has served as publica- uted Systems, IEEE Transactions on Computers, IEEE Transactions on
tion chairs, publicity chairs, section chairs, and TPC member for numer- Computer-Aided Design of Integrated Circuits and Systems, IEEE Trans-
ous conferences. He has been an associate editor of the Journal of actions on Very Large Scale Integration (VLSI) Systems, IEEE Transac-
Circuits, Systems, and Computers, and serves as a guest editor for sev- tions on Circuits and Systems, IEEE Transactions on Smart Grid, IEEE
eral special issues of the ACM Transactions on Cyber-Physical Systems, Transactions on Dependable and Secure Computing, IEEE Transactions
IET Cyber-Physical Systems: Theory & Applications, and Elsevier Jour- on Aerospace and Electronic Systems, IEEE TR, IEEE Transactions on
nal of Systems Architecture: Embedded Software Design. He is a mem- Sustainable Computing, IEEE Communications Surveys and Tutorials,
ber of the IEEE. etc. He has been a regional editor of the Journal of Circuits, Systems, and
Computers since 2012. He served as a guest editor for several special sec-
tions of the IEEE Transactions on Industrial Informatics, ACM Transactions
Gongxuan Zhang (SM’12) received the BS on Embedded Computing Systems, and ACM Transactions on Cyber-
degree in electronic computer from Tianjin Univer- Physical Systems. He is a senior member of the IEEE.
sity, in 1983, and the MS and PhD degrees in
computer application from the Nanjing University
of Science and Technology, in 1991 and 2005, Shiyan Hu (SM’10) received the PhD degree in
respectively. He was a senior visiting scholar with computer engineering from Texas A&M University,
the Royal Melbourne Institute of Technology from in 2008. He is currently a professor and chair in
2001 to 2002 and with the University of Notre cyber-physical system security with the University
Dame from 2017 to 2017. Since 1991, he has of Southampton. His research interests include
been with the Nanjing University of Science and cyber-physical systems and cyber-physical sys-
Technology, where he is currently a professor tem security, where he has published more than
with the School of Computer Science and Engineering. His current 100 refereed papers. He is an ACM distinguished
research interests include multicore and parallel processing and distrib- speaker, an IEEE Systems Council distinguished
uted computing. He is a senior member of the IEEE. lecturer, an IEEE Computer Society distinguished
visitor, a recipient of the 2017 IEEE Computer
Society TCSC Middle Career Researcher Award, 2014 U.S. National
Science Foundation (NSF) CAREER Award, and 2009 ACM SIGDA
Richard Newton DAC scholarship. His publications have received a few
distinctions, which includes the 2018 IEEE Systems Journal Best Paper
Award, 2017 Keynote Paper in IEEE Transactions on Computer-Aided
Design, and the Front Cover in IEEE Transactions on Nanobioscience in
March 2014. He is the chair for IEEE Technical Committee on Cyber-
Physical Systems. He is the editor-in-chief of the IET Cyber-Physical Sys-
tems: Theory & Applications. He serves as an associate editor of the IEEE
Transactions on Computer-Aided Design, IEEE Transactions on Industrial
Informatics, IEEE Transactions on Circuits and Systems, ACM Transac-
tions on Design Automation for Electronic Systems, and ACM Transac-
tions on Cyber-Physical Systems. He has served as a guest editor for
eight IEEE/ACM Journals such as the Proceedings of the IEEE and IEEE
Transactions on Computers. He has held chair positions in many IEEE/
ACM conferences. He is a fellow of the IET and British Computer Society,
and a senior member of the IEEE.

" For more information on this or any other computing topic,

please visit our Digital Library at www.computer.org/csdl.