Int. J. Communications, Network and System Sciences, 2011, 4, 803-811
doi:10.4236/ijcns.2011.432098 Published Online December 2011 (http://www.SciRP.org/journal/ijcns)
IaaS Public Cloud Computing Platform Scheduling
Model and Optimization Analysis
Aobing Sun1,2,3, Tongkai Ji1,2,3, Qiang Yue1,2,3, Feiya Xiong3
1Guangdong Electronic Industrial Institute, Dongguan, China
2Institute of Computing Technology of Chinese Science Academy, Beijing, China
3Sino-Cloud Science and Technology Stock Company Ltd., Dongguan, China
Received August 30, 2011; revised September 10, 2011; accepted October 14, 2011
Abstract
IaaS (Infrastructure as a Platform) public cloud is one mainstream service mode for public cloud computing.
The design aim of one IaaS public cloud is to enlarge the hardware-usage of whole platform, optimize the
virtual machine deployment and enhance the accept rate of service demand. In this paper we create one ser-
vice model for IaaS public cloud, and based on the waiting-line theory to optimize the service model, the
queue length and the configuration of scheduling server. And create one demand-vector based scheduling
model, to filter the available host machine according to the match of demand and metadata of available re-
source. The scheduling model can be bonded with the virtual machine motion to reallocate the resources to
guarantee the available rate of the whole platform. The feasibility of the algorithm is verified on our own
IaaS public cloud computing platform.
Keywords: Cloud Computing, IaaS, Scheduling Model, Optimization Analysis
1. Introduction
Cloud Computing Technology is developed from virtu-
alization, utility computing, IaaS (Infrastructure as a
Service), PaaS (Platform as a Service), SaaS (Software
as a Service) and etc. [1]. It puts forward one new IT
business model, i.e. the users can acquire IT services
through Internet with o n-demand and expandable means.
The cloud computing platform utilizes the high-speed
Internet to deliver the computing, storage, software and
services which are distributed all over the world, to the
terminal users and make them to use the resources as
electricity. The cloud computing technology brings us a
new service mode to serve the users with data, applica-
tion and IT r e sources through network [2].
Cloud computing technology is also one methodology
for infrastructure, i.e. the cloud computing platform inte-
grates the mass computing resources to compose one
resource pool and serve the users dynamically with vir-
tualized resources including computing, storage and ser-
vice. To one user of cloud computing platform, almost
everything as software, hardware, data and information
service all can be rent from the cloud. The cloud com-
puting platform can be subdivided into three layers
shown as Figure 1 [3].
SaaS (Software as a Service) i.e. the software is de-
livered through Web browsers as a service of cloud
computing platform, so the users can rent the software on
demand. SaaS of cloud computing includes SaaS soft-
ware and trusteed applications, e.g. Saleforce is one fa-
mous SaaS provider; it delivers ERP, SCM, CRM soft-
ware and etc. through Internet with SaaS mode [3].
PaaS (Platform as a Service) provides one platform
for the users and developers with application develop-
ment, test and deployment, e.g. one SaaS application.
The platform includes database, middleware and devel-
opment tools, and all services can be composed through
Internet. For example, the Google Map platform and
APP platform all are the PaaS cloud platform [2].
IaaS (Infrastructure as a Service) is to provide the
hardware infrastructure as servers, storage and hardware
through Internet. The IaaS platform is created based on
virtualization technology as server and storage virtual-
ization, so virtualization, cluster and dynamic configura-
tion software are also includes IaaS. e.g. EC2 of Amazon
is one famous IaaS platform of cloud computing tech-
nology [1].
The cloud computing platforms own three types:

804 A. B. SUN ET AL.
Figure 1. The layered structure of cloud computing plat-
form.
Public Cloud serves the users that distributed all over
the world across the border of enterprises and areas.
Usually the public cloud platform is large-scale and
composed by a few data centers in different area to pro-
vide IaaS, PaaS or SaaS service. e.g. Amazon EC2 is the
IaaS public cloud, Google APP and Apple AppStore is
the PaaS public cloud [3]. Public cloud serves the gen-
eral users with on-demand mode, so the small enterprise
users can create their IT business systems with low-cost.
But the supervision of public cloud is very difficult, e.g.
the resources of EC2 are used for spam mails, hack at-
tack and Trojan attack [4].
Private Cloud only serves for one company or or-
ganization. Generally private cloud is composed by IT
infrastructure of one enterprise. It contains their data
center and all other devices, hardware and software
linked in Internet. The private cloud is managed by the
IT fellow and with high-level security. Private cloud de-
mands the entire control of resources, and react the users
with different priorities. So the users can have specific
demands to resources. But generally, the pubic cloud
looks the users with same priorities. The widely used
private cloud includes VCloud, VSpher e of VMware and
XEN Cloud of Citrix [5].
Mixed Cloud owns the properties of public cloud and
private cloud. It connects the resources of private clouds
including its data, application and serv ice through public
cloud, e.g. private cloud connects into one public cloud
and provide one access interface through one agent
server. So it can guarantee the security of private cloud
and support the permitted resources can be exposed into
Internet. OpenNebula is one famous mixed cloud plat-
form [3,6].
In this paper we describe our IaaS public cloud plat-
form. The rest of the paper is organized as follow: Sec-
tion 2 relates the service model of IaaS public cloud.
Section 3 states our scheduling model for IaaS public
cloud and its optimization means. Section 4 gives some
experimental results. And Section 5 draws one conclu-
sion and gives out f ut ure works.
2. Model of IAAS Public Cloud
IaaS public cloud is one important application mode of
current cloud computing technology. With the appear-
ance of Amazon EC2, more and more platforms come
out to provide computing and storage resources. The aim
of the platforms is to provide the users on demand with
the virtual machines for ordered CPU frequency, quan-
tity of core, storage space and memory size [4].
2.1. Element of IaaS Public Cloud
As shown in Figure 2, logically one IaaS public cloud
owns three main elements as follow:
Cloud Administration Center is the access interface to
Internet and also the management, scheduling and moni-
toring center of the resources within the cloud. The ad-
ministration center of one IaaS public cloud accepts the
resources request from the Internet users and create the
demanded resources, e.g. virtual machine and storage
resources, and configure them, then return the resources
to the users.
Cloud computing Resources Center is composed by
the physical computing resources. To one IaaS platform,
the physical resources will be used as the host machines
to be administrated by the cloud administration center.
The scheduling server will select the optimal resources
according to the user demands to create virtual machines.
In general, multiple cloud computing resource centers
access the administration center with agent servers which
can also be used to support the monitoring and schedul-
ing of the computing resources.
Cloud Storage Resources Center is composed logi-
cally by the physical storage resources. To one IaaS
platform, virtual machine template, images and snap-
shots are also stored in the storage center which is ad-
ministrated with network storage systems as NFS, S3,
ISCSI and etc. The virtual machine image of users is
transferred to one specific physical machine from the
storage center and then is loaded into it. To the platform,
the physical and virtual machines are loosely coupled.
And it also is the difference of public and private cloud.
Service Workflow
We will use two operation flows to analyze the sched-
uling flow of cloud platform. The flow of user request to
resources is as follow:
1) The registered users access the portal server and
request the virtual machines with the parameters includ-
ing quantity of core, frequency, memory, storage space,
OS and etc.

A. B. SUN ET AL.
805
Figure 2. The elements of IaaS public cloud.
9) The monitoring server will renew the information
within the metadata database, to guarantee its correctness,
and to improve the efficiency of sch eduling operations.
2) Portal Server sends the request to the scheduling
server.
3) The scheduling server searches the physical ma-
chines to find the host to create the virtual machine ac-
cording to the metadata of physical machine, which re-
cords the operation and configuration details.
2.2. The Service Model of IaaS Public Cloud
4) The scheduling server chooses one optimal server
and then sends the creating command of virtual machine
to its agent server.
According to the service flow, we can abstract one IaaS
Public cloud as the model as shown in Figure 3. The
model including three flows as follow:
5) The scheduling sever chooses the virtual machine
template form the stored templates within cloud storage
administration center, and sends one request for the tem-
plate to the agent server.
1) The scheduling severs of cloud administration cen-
ter, picks out the request R with the highest priority. The
scheduling server then judges whether R can be meted
according to the parameters of R, as CPU frequency, core
quantity, band-width, storage, and disk sp ace. If R can be
met, then jump to (2). Or then judges if R can be met
through the VM (Virtual Machine) motion and then re-
lease enough resources; If one motion can release
enough resources, then jump to (2). Or then quit directly
and report to the user that the requested resources cannot
be met.
6) The requested virtual machine image will be sent
(or mapped) to the physical sever based on the template,
the scheduling sever will start the virtual machine if the
image is loaded successfu lly. If something is wrong dur-
ing (4)-(6), the scheduling server will select new virtual
machine.
7) If the virtual machine starts successfully, the user
can access the virtual machine through RDP, VNC, ICA
or SSH. 2) If the requested resources can be met, the schedul-
ing server then choose the VM template T (if create one
new VM) or one VM image I (if use existing VM) cor-
responding to R.
The agent server can monitor the resources registered
within the computing or storage resources management
center. It will renew the metadata within the metadata
database to guarantee the correctness of scheduling serv-
er to resources. The renewing of metadata follows 2
steps:
3) The scheduling server sends I to corresponding
physical machine and create VM instance V.
There are three important problems to the model:
1) How the request-queue length is determined, and
the priority of request R can be adjusted, to guarantee all
the request can be reacted quickly.
8) The monitor server will send the resource-informa-
tion renew request, the scheduling server will send the
request to the agent servers. If the agent severs acquire
the information then send them to the monitoring server. 2) How the request R can be parsed and then to select
the optimal resource to serve users.

A. B. SUN ET AL.
806
Figure 3. IaaS public cloud service model.
3) If VM motion operations are demanded, and how to
guarantee the motion cost and the affection to the other
VMs are minimized. To one platform the motion can
affect the QoS of whole platform, so the motion of VM
will be executed on ly when other VMs will not be disor-
dered.
3. Scheduling Model of IAAS Public Cloud
According to the service model, we can quantize the pa-
rameters and analyze the throughout of one cloud plat-
form, and then analyze and optimize the model [6].
3.1. Queue Model
According to the waiting-line th eory, as shown in Figure
4, the request and react processing is one waiting-line
system, the input of one waiting-line system is the re-
quest and the service counter is the scheduling server,
and the output is the requested resources. One user re-
quest queue is .

123
,,,,
n
RRRR R
We assume the request come as Poisson’s ratio within
one IaaS public cloud, and the service time is as expo-
nential distribution. λ is the count of coming user request
averagely in one unit of time. μ is the service efficiency
(the ability of service counter). ρ = λ/μ is the ration that
the request can be met within one unit of time, i.e. the
service success rate. WS is the time that one service will
wait in the system, which includes the waiting time and
service time. Wq is the average waiting time for user re-
quest. If there only is one counter (i.e. one scheduling
server), so


1
S
W



,

q
W

  . So there are
two means to increase the user request reaction speed:
1) decrease the count of user request sent to schedul-
ing server.
2) Increase the processing speed of scheduling server.
So we can increase scheduling server within one IaaS
public cloud. When there are several scheduling server,
the results are as follow equations.


0
2
1
!1
k
S
k
W
k

T




 (1)


0
2
!1
k
q
k
W
k

T


 (2)

1
1
00
11
!!1
nk
k
n
Tnk








 



 


  


 (3)
Multiple-queue and multiple-counter can be seen as
several single-queue and single-counter. So based on the
analysis, to control the reaction speed of system, the
maximum of queue length will be fixed. When the wait-
ing request surpass it, the requests out of the queue will
be rejected. So the maximum reaction speed of user re-
quest is the whole queue processing time [7].
3.2. Model Analysis
The set of physical machine within one cloud in P,


123
,,,,
n
PPPP P. n is the count of physical mach ine
within P. All the parameters are as shown in Table 1.

A. B. SUN ET AL.
807
Table 1. Notation of quantized scheduling model.
Notation Presentation
Pi Anyone physical machine within P
Ci Sum of allocable CPU core of Pi
Fi Sum of allocable CPU frequency of Pi
Mi Sum of allocable memory of Pi
Bi Sum of allocable network bandwidth of Pi
Di Sum of allocable dis k space of Pi
Vi VM set running on Pi


i
F
V Sum of used CPU frequency of Vi


i
CV Sum of used CPU core of Vi

i
M
V Sum of used memory of Vi


i
D
V S um of used disk space of Vi

i
B
V Sum of used network bandwidth of Vi
To user request R, the resource allocatio n must follow
the rules which are also the necessary rules of one IaaS
public cloud.
1) To one single VM, anyone of the allocated resour ce
to vij of Vi (as frequency, core quantity, disk space and
bandwidth) will be less than the total resources of Pi, i.e.


ij i
f
vF

, and


ij i
cv C

, and , and

ij i
mv M


dv D
ij i

, and


bv B
ij i

.


ij
f
v

,

ij
cv ,


ij
mv ,


ij
dv and


ij
bv is the allocated frequency, core,
memory, disk and bandwidth to VM vij.
2) The sum of the allocated resources to the virtual
machines within VM set Vi will be less than the total of
physical machine Pi, i.e. i

i
F
VF, and


ii
CV C

,
and


ii
M
VM, and


ii
DV D, and

ii
BV B

.
As shown in Figure 5, assuming the user request Ri
can be parsed into CPU frequency request RFi, CPU core
request RCi, disk request RDi, memory request RMi, net-
work bandwidth request RBi. The scheduling server
firstly goes through the physical machines within the
metadata record, and to find the physical-machine set
that can meet the VM request. Then sort the physical
machines according to its usage. The VM will be created
on the physical machine with the lowest usage. The usage
is one powered remark including CPU frequency, memory
and bandwidth usage. Generally the CPU usage can be
Figure 4. Waiting-line model of IaaS public cloud re que st.

 




        
FindPhysical Computer(Request, Computer Set)
{empty(PM); //Emptytheresultsetofphysicalmachine
for
{if
//Remained resource ofphysical machine canmee
i
iiii ii
iiiiiiii i
RP
P
PP
RFFF VRCCC V
RMMFMVRDDDV RBBBV

 
 


∨





1
1
11
tthe request
theninsertPM; //insert into the queue
}
forPM//To anyone physical machine inPM
{if
//Ifthe usage ofismore than
then; //andexchangedwithinthe set
}//Sortthe physical machine wit
ii
i
ii
ii
ii i
PP
P
UP UP
PP
PPP P






∨
hin PMaccording toUsage
return PM;
}
Figure 5. Algorithm to find physical machine to host one VM .

A. B. SUN ET AL.
808
used as the main indicator of total usage.
One public cloud platform can release resources
through VM motion to meet one request. Because one
VM motion will decrease the QoS of VMs within same
physical machine, the platform will decrease the possible
VM motions and use at the most one motion to meet the
resources release demand. When the physical machine
within one physical machine set all cannot meet the de-
mand, the scheduling server firstly find two physical-
machine with the lowest usage, and attempt to move the
VM with the lowest usage to another physical machine to
release resources. To the physical machine, if one step
motion cannot release enough resources, the user request
will be refused. The algorithm to release resources is as
shown in Figure 6.
4. Experiment
G-Cloud v3.0 is one IaaS public cloud computing plat-
form developed by GDEII (Guangdong Electronic In-
dustrial Institute). In this paper, we use one group of ex-
periment to verify our scheduling model. The result is as
shown in Table 2.
The experimental platform owns 100 physical ma-
chines as the host, and one host can create 8 VM at most.
We change the length of request queue and put forward
the VM creation request to surpass the length to test our
algorithm. We can see from Table 2 that with the im-
proved multi-scheduler means the minimum reaction
time will not be affected by the length of the request
queue with FIFO (First In First Out) Mode. But the av-
erage reaction time and the maximum reaction time will
be enlarged with the increase of the queue length as
shown in Figure 7(a), especially when the queue length
surpasses 40 s. And the maximum waiting time will sur-
pass 120 s which is over the user-enduring time. So with-
in our platform, the request queue length is 40. Con-
trasted with the general means, only with one scheduler

  
1
1
MotionPhysical Computer(Request, Computer Set)
{em p ty(PM); // Emptytheresultset
for//Toanyonephysicalmachinewithin
{if
//Ifthe usage ofis less thanthreshold
theninsert; //Insertinto
//F
i
i
i
ii
RP
PM
PP P
UP T
PT
PPM PPM


∨





2
2
indthe VM whose usage isless thanThreshold
}
for PM
//To any physicalmachineofPM
{if
//Ifthe usage ofis less thanthreshold
theninsert VM; //Insertinto VM
//Find VMs thataremovable to aim physical ma
ij ii
ij
ij
ij ij
VV P
UV T
VT
VV


∨∨



     
     
1
1
chine
}
forVM//To any virtual machine ofVM
{if(Motionabled,is True)
{//If ismovabletoonephysicalmachine
PreMotion,//Pre-move one VM logically
if
ij
ij i
ij
ij i
iiii ii
iiiii ii
V
VP
V
VP
RFFF VRCCC V
RMMFMVRDDDV RB



 
 


∨
  

1
//Ifrequested resources canbe met aftermotion
{Motion,//Move the VM really
returntrue; //Return Success
}
}
return false;//Cannot find motionaim,returnfalse
}
ii
ij i
BBV
VP


Figure 6. Algorithm to find VM for motion.

A. B. SUN ET AL.
809
Table 2. Queue length and reaction tims(s) with multi-
scheduler.
Queue
Length Minimum Reaction Time Average Reaction Time Maximum Reaction Time
10 1.4 12.3 26.4
20 1.35 17.8 30.21
30 1.41 19.1 39.11
40 1.54 26.7 42.11
50 1.3 31.8 55.11
60 2.1 46.6 65.34
70 1.3 50.2 120.12
80 1.5 60.3 165.6
90 2.1 75.9 170.4
100 2.3 99.0 220.1
and without queue length adjust algorithm the average
reaction time will increase with very high speed and
surpass 40 s with only 20 jobs in queue as shown in Fig-
ure 7(b). So it is very important to constru ct more sched-
ulers and the queue length should have one maximum
and can adjust dynamically.
When the VM capacity of the whole platform sur-
passes 80% and if we create new VM or reconfigure one
VM, the resources will not be enough to create it. So the
platform will move some VM to release resource to meet
the demand. The VM motion will be affected by the VM
image, memory and storage size, and the backup, sched-
uling and motion algorithm. Out platform adopts Hot
Motion means (related within Figures 5 and 6), i.e. the
VM is moved when it is running . But the data coherence
cannot be kept easily. After the VM image is synchro-
nized, the VM will stop service for several seconds to
synchronize the runtime memory. After image synchro-
nization, the original VM will be closed and the new VM
(a)
(b)
F
igure 7. Queue length and reaction time. (a) Improve d means wi th multi -sche duler; ( b) Contra st of general and impr oved means.

A. B. SUN ET AL.
810
will start service. Hot motion will affect the online users
and the QoS. But it can keep service for online users
after hot motion.
Contrasted with our algorithm, there are the cold mo-
tion and clone motion means. Cold motion means the
VM is moved after it is closed. Clone Motion, i.e. the
VM motion will only move the image files and the run-
time memory file will be abandoned. So the QoS will be
affected. And the Clone motion only is used when the
VM will have no data renew. And the online users will
be disconnect ed du r i n g clone motion.
We usually use one time-cost to quantity the motion
cost for VM motion. It includes the cost for VM image
and memory files transferring, reconfiguration and restart
VM. But the service stop time is the main scale for the
VM motion of one IaaS public cloud platform. We can
see from Figure 8 that the hot motion will cost more
time than cold motion and clone motion. Because it need
more time to avoid the service delay time and use more
time to synchronize the VMs. But the hot motion will
cause the minimum delay time than cold motion and
clone motion as shown in Figure 9.
Figure 8. Total time cost for different motion mode.
Figure 9. Service delay time for different motion mode.

A. B. SUN ET AL.
811
] K. Chen and W. M. Zheng, “Cloud Computing: System
493
5. Conclusions and Future Works
IaaS public cloud aims to provide available VMs for
Internet users. In this paper, we summarize the IaaS pub-
lic cloud model, and analyze the service flow according
to the waiting-line theory. And aiming to maximize the
platform usage and the performance of single VM, we
give out one filtering algorithm based on user request to
find optimal resour ce for user VM req uest [8]. Th e cloud
administration center renews the metadata on time to
support the virtual machine motion and physical machine
scheduling [9]. The algorithm is verified on our G-Cloud
platform of GDEII, which improves the QoS of the
whole plat fo rm.
6. Acknowledgements
This work is partially supported by the Strategic Coop-
eration Project of Guangdong Province and Chinese
Science Academy Grant # 2009A0091100002 and 2010-
A090100004; Supported by Guangdong and Hong Kong
invited bidding special for Dongguan Grant # 2011-
20510101, 201120510106 and 201120510104; And sup-
ported by Dongguan Major Science and Technology
Special Project Grant # 2009215102001.
7. References
[1
Instance and Current State,” Journal of Software, Vol. 20,
No. 5, 2009, pp. 1337-1348.
doi:10.3724/SP.J.1001.2009.03
lassification Research
Based on
nd S. Venugopal, “Market-Orien-
a Science of the Web,”
[2] Z. W. Xu, H. M. Liao, et al., “The C
of Network Computing System,” Journal of Computing
Machine, Vol. 18, No. 9, 2008, pp. 1509-1515.
[3] G. W. Zhang, R. He an d Y. Liu, “The Evol ution
Cloud Model,” Journal of Computing Machine, Vol. 7,
2008, pp. 1233-1239.
[4] R. Buyya, Y. S. Chee a
ted Cloud Computing: Visio n, Hype, and Reality for Del iv-
erring,” 10th IEEE International Conference on High Per-
formance Computing and Communications, Dalian, 25-27
September 2008, pp. 5-13.
[5] T. Berners-Lee, et al., “Creating
Science, Vol. 313, No. 5788, 2006, pp. 769-771.
doi:10.1126/science.1126902
[6] A. Kemal Delic, “On Dependability of Corporate Grids,”
Ubiquity, Vol. 6, No. 45, 2005, p. 1.
doi:10.1145/1113164.1113162
[7] K. Delic and M. Walker, “Architecting Enterprise Grids:
riented Science,” Science, Vol. 308,
Possible Inflection Points”, IADIS International Confer-
ence on Applied Computing, Salamanca, 18-20 February
2007, pp. 113-121.
[8] I. Foster, “Service-O
No. 6, 2005, pp. 814-817. doi:10.1126/science.1110411
[9] E. Hand, “Head in the Clouds,” Nature, Vol. 449, No. 24,
2007, pp. 963-970. doi:10.1038/449963a