Tuijin Jishu/Journal of Propulsion Technology
ISSN: 1001-4055
Vol. 44 No. 5 (2023)
__________________________________________________________________________________________
2-D Scaling in Cloud computing
[1]
Aruna Mailavaram, [2] B. Padmaja Rani,
[1]
Asst. Prof. in CSE, Madanapalle Institute of Technology & Science,A.P.
[2]
Prof. in CSE, JNTUHCEH, Hyderabad.
Abstract: Two Dimensional scaling is increasing the resources of the cloud as per the requirements. In this
paper we are discussing about the scaling in both vertical and horizontal as to fulfill the needs of the
customers. Vertical scaling is increasing the specifications of existing infrastructure by adding or replacing
CPU, HDD or other components, whereas horizontal scaling is adding new resources to the existing pool of
resources. Increasing the resources in two-dimensional is known as Diagonal Scaling.
Key Words: Deadline, scaling, cloud computing, performance, vertical scaling and horizontal scaling.
Introduction
Cloud Computing[6] is a huge source of resources where different organizations utilize them for their
work. Since all the resources are globally spread through the internet, remotely many services can be accessed.
Since the customers are not limited, day by day customers are increasing cloud has to be designed with no
issues. But still we have few issues in cloud computing like security, scalability, performance, latency etc. To
design a scalable[1] cloud system to improve performance and making all the tasks to reach within deadline
time we are proposing a method, where all the tasks are reaching the deadlines.
The above figure shows the architecture of the cloud, how the cloudlets[3] are being processed using
cloud sim tool, this architecture provides the environment of cloud. The scheduling policies[4] will be mounted
on scheduling policy in the architecture[5]. Here in this paper we are designing horizontal and vertical scaling
algorithms and compared with each other. We are trying to improve the performance of the system as well as
the tasks should reach the deadline. So two parameters we are focusing on. In my next upcoming paper, we will
focus on the combination of both vertical and horizontal scaling algorithms.
3806
Tuijin Jishu/Journal of Propulsion Technology
ISSN: 1001-4055
Vol. 44 No. 5 (2023)
__________________________________________________________________________________________
Design and Implementation:
Here we are implementing both horizontal and vertical scaling[12] separately and compared with the
previous algorithms and found better results in horizontal scaling[13]. Initially we are taking a data set of VMs,
task lengths and deadlines as input and given to broker function for accessing all the tasks are arranged in
ascending order and VMs in descending order. Allotting tasks to the VMs for processing if any leftover tasks are
there we are implementing horizontal scaling[12].
Horizontal Scaling : Arrange all the VMs in descending order, and all tasks in ascending order
associated with their deadlines according to the task length[9]. Here in this algorithm we are checking
overloaded condition and under loaded condition of each VM. We are utilizing 80 percentage of each MV
capacity in order to avoid overloaded condition. For example if we consider 720 MIPS capacity VM then we are
utilizing 576 MIPS of its capacity. And we are seeing that tasks not reaching the deadline is less at the same
time improving the performance of the system.
Since this is horizontal scaling we can add resources to the pool of existing resources[8]. In the given
data set we have taken ten VMs and arranged them in descending order. Two VMs with high capacity are kept
in spare for further use to handle high length tasks. So we are starting from 720 MIPS Virtual Machine, and
assigning the tasks until it reaches 80 percentage of VM capacity[3], then task is being hand over to next VM,
until all VMs get engaged. Check for the left over tasks, assign the tasks to the VMs which we kept in spare for
further use to handle high length tasks. If further any tasks ae still left over then we have to implement
horizontal scaling if the percentage of unprocessed tasks[10] are less than 10 percentage then increase VMs with
10 percentage, if else the unprocessed tasks are between 10 to 25 percentage increase the VMs by 25 percentage
else calculate Average Resource Utilization Ratio (ARUR).
Problem formulation :
Capacity of VM is CVM = p*q+BW{j,k}
Where p is the processing speed and q is he umber cpu and BW is the band width between j and k virtual
machines.
Now we find out the load at the data center i.e., DC =∑i=1mCVMj
Load information at the VM can be found by LVM = number of task *∑task length/p*q.
So, total load at Data Center is TL VM =∑j=1mLVM.
Expected Execution Time at Virtual Machine is calculated as
EET = Task Length/p*q.
Makespan Time = Max( ∑mi=1ET.
Algorithm for Horizontal Scaling :
[Link] n tasks T1,T2,T3,…Tn.
[Link] the tasks in ascending order.
3. Generate m number of virtual machines and sort them in descending order.
4. Consider VMs from m*0.2 to m-1, higher capacity VMs are kept in UVM list in increasing order to handle
high length tasks.
5. For loop for T i from 0 to n-1 and for all virtual machines Rj belongs to m-1.
6. If(Tasks[i]!=Ø) then
7. For loop Tasks[i] -> VM[j], i++ until
Total _time TT[j]<(VM[j]*0.8)
[Link] allot tasks to next VM i.e., VM[j+1]
[Link] step 7 and 8 until all VMs are engaged.
[Link](Tasks[i]==Ø) then calculate makespan
[Link] If
While (UVM !=0)
Tasks[i+1] -> VM[j] until UVM ==0
12. Else calculate the left over tasks
3807
Tuijin Jishu/Journal of Propulsion Technology
ISSN: 1001-4055
Vol. 44 No. 5 (2023)
__________________________________________________________________________________________
i. If (UT>0 && UT< m*0.1) increase VMs by 10 percentage
ii. Else If (UT>m*0.1) Increase VMs by 20 percentage
iii. Else UT==0 calculate ARUR.
13. Repeat step 11 and step 12 until all tasks are completed.
Vertical Scaling : In vertical scaling[14] initially we are using the same technique as we did in
horizontal scaling to allot the tasks to the VMs. Once the left over tasks are available, then allot them to the VM
which are in idle state, but these left over tasks has to wait until previous tasks allotted to the virtual machine is
completed and then allot task to the VM so here we have to increase the VM capacity based on the deadline
achievement.
Increasing the power of the system is vertical scaling unlike horizontal scaling the system power in
terms of cpu, RAM[11] we can increase in vertical scaling, many public clouds can also allow to increase the
capacity of the system power (example AWS). But the Virtual machine has to stop running we can increase its
power and start running with the new power to handle the task to complete within the deadline.
In this scaling technique we are increasing the capacity of the VM based on the deadline, a new VM is
generated to reach the task within the deadline and still increase the capacity of the VM by 25 percentage to
overcome the overloaded condition and round off the capacity of the VM.
Problem formulation :
New_VM=Task[i]/Deadline[i]
To avoid overload condition we are increasing the VM capacity by 25 percentage.
VMIncrease = New_VM*25/100
New_VM = New_VM + VMIncrease
Algorithm : Vertical Scaling
Input : VMs, Tasks and Deadline.
Output : All tasks are being processed.
[Link] n tasks T1,T2,T3,…Tn.
[Link] the tasks in ascending order.
3. Generate m number of virtual machines and sort them in descending order.
4. Consider VMs from m*0.2 to m-1, higher capacity VMs are kept in UVM list in increasing order to handle
high length tasks.
5. For loop for T i from 0 to n-1 and for all virtual machines Rj belongs to m-1.
6. If(Tasks[i]!=Ø) then
7. For loop Tasks[i] -> VM[j], i++ until
Total _time TT[j]<(VM[j]*0.8)
[Link] allot tasks to next VM i.e., VM[j+1]
[Link] step 7 and 8 until all VMs are engaged.
[Link](Tasks[i]==Ø) then calculate makespan
[Link] If
While (UVM !=0)
Tasks[i+1] -> VM[j] until UVM ==0
[Link] calculate the number of left over tasks.
13. If UT!=Ø then
14. Generate new VM based on the deadline for each task as New_VM=tasks[i]/Deadline[i]
15. To avoid overload condition increase the new VM by 25 percentage.
Synthetic Data :
Considering a dataset for both the algorithms.
3808
Tuijin Jishu/Journal of Propulsion Technology
ISSN: 1001-4055
Vol. 44 No. 5 (2023)
__________________________________________________________________________________________
Task ID Task Length(MI) Deadline of Task VMs (MIPS)
{ms)
17 98049 1050 760
18 107218 300 740
19 147281 1150 720
15 153285 800 700
7 182561 2000 680
16 205633 600 660
9 215744 1300 640
13 222784 550 620
10 253197 1250 600
14 253745 1000 580
4 325750 420 560
3 325800 410
5 325970 700
6 333911 660
11 334013 450
2 339760 920
12 344630 400
0 381771 400
1 392397 745
8 396156 300
After implementing Horizontal and vertical scaling we are getting the result as follows:
Results
Horizontal Scaling:
Task ID Task Length Deadline of Task Execution Time
17 98049 1050 136
18 107218 300 284
19 147281 1150 488
15 153285 800 218
7 182561 2000 478
16 205633 600 302
9 215744 1300 326
13 222784 550 348
10 253197 1250 408
14 253745 1000 422
4 325750 420 561
3 325800 410 577
5 325970 700 440
6 333911 660 439
11 334013 450 428
2 339760 920 424
12 344630 400 420
0 381771 400 454
1 392397 745 456
8 396156 300 450
3809
Tuijin Jishu/Journal of Propulsion Technology
ISSN: 1001-4055
Vol. 44 No. 5 (2023)
__________________________________________________________________________________________
There are five tasks which are not meeting the deadline[7] Task ID 4,3,12,0,8.
Vertical Scaling:
Task ID Task Length Deadline of Task Execution Time
17 98049 1050 136
18 107218 300 284
19 147281 1150 488
15 153285 800 218
7 182561 2000 478
16 205633 600 302
9 215744 1300 326
13 222784 550 348
10 253197 1250 408
14 253745 1000 422
4 325750 420 420
3 325800 410 410
5 325970 700 756
6 333911 660 978
11 334013 450 946
2 339760 920 878
12 344630 400 848
0 381771 400 863
1 392397 745 1038
8 396156 300 1038
In vertical scaling there are total 8 tasks which are not reaching the deadlines[2] they are Task ID
5,6,11,2,12,0,1,8.
ARUR(Average Resource Utilization Ratio) :
Average Resource Utilization Ration(ARUR) for Vertical Scaling
ARUR = (Mean Time / Make Span )* 100
Mean Time = ∑ Total Time taken by resource VMj to finish all the job / No. of resources.
Parameters in Horizontal Scaling :
Parameters Values
No. of Overloaded Machines 2
No. of Under loaded Machines 0
Tasks not reaching deadline 5
Makespan 577ms
ARUR 75.65
Parameters in Vertical Scaling :
Parameters Values
No. of Overloaded Machines 8
No. of Under loaded Machines 0
Tasks not reaching deadline 8
Makespan 1038 ms
3810
Tuijin Jishu/Journal of Propulsion Technology
ISSN: 1001-4055
Vol. 44 No. 5 (2023)
__________________________________________________________________________________________
ARUR 83.01
Comparison (ms) :
Horizontal Scaling 577
Vertical Scaling 1038
84
82
80
78
76
74 Series1
72
70
Horizontal Vertical Scaling
Scaling
Number of Overloaded machines :
Horizontal Scaling 2
Vertical Scaling 8
84
82
80
78
76
74 Series1
72
70
Horizontal Vertical
Scaling Scaling
Number of tasks not reaching deadline :
Horizontal Scaling 5
Vertical Scaling 8
84
82
80
78
76
74 Series1
72
70
Horizontal Vertical
Scaling Scaling
ARUR :
Horizontal Scaling 75.65
Vertical Scaling 83.01
3811
Tuijin Jishu/Journal of Propulsion Technology
ISSN: 1001-4055
Vol. 44 No. 5 (2023)
__________________________________________________________________________________________
84
82
80
78
76
74 Series1
72
70
Horizontal Vertical
Scaling Scaling
Conclusion
We have discussed about both vertical and horizontal scaling in cloud computing separately, scale out
is horizontal scaling on x-axis and scale up vertical scaling on y-axis. In my next coming papers we will
combine horizontal and vertical scaling. But according to this paper we concluded that in horizontal scaling we
were able to achieve more number of tasks meeting the deadlinewhen compared with vertical scaling.
References
[1] Dynamic load balancing algorithm for balancing the workload among virtual machine in cloud
computing Mohit Kumara,*,[Link].[Link]
[2] Deadline constrained based dynamic load balancing algorithm with elasticity in cloud environment.
[Link]
[3] Elastic and flexible deadline constraint load Balancing algorithm for Cloud
Computing.[Link]
[4] Dynamic load balancing algorithm to minimize the makespan time and utilize the resources effectively
in cloud environment.
[Link]
[5] Load balancing algorithm to minimize the make span time in cloud environment. ISSN1746-
7233,England,UK World Journal of Modelling and Simulation. Vol.14(2018)No.4,pp.276-288.
[6] A comprehensive survey for scheduling techniques in cloud computing.
[Link]
[7] Dynamic Auto-scaling and Scheduling of Deadline ConstrainedService Workloads on IaaS Clouds.
10.1016/[Link].2016.05.011
[8] Optimized Task Scheduling and ResourceAllocation on Cloud Computing EnvironmentUsing
Improved Differential Evolution Algorithm. [Link]
[9] Heuristic-based load-balancing algorithm for IaaS cloud.[Link]
[10] Load balancing in cloud computing: A big picture. [Link]
[11] A Priority Based Job Scheduling Algorithm Using IBA and EASY Algorithm for
Cloud Metaschedular[Link]
[12] Horizontal and Vertical Scaling of Container-Based Applications Using Reinforcement Learning.
[Link]
[13] Vertical Scaling of Virtual Machines In Cloud Environment.
[Link]
[14] Vertical/Horizontal Resource Scaling Mechanism for Federated Clouds.
[Link]
3812
