Energy-Efficient Task Scheduling in Cloud Environment

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 04 | Apr 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 2679
Energy-Efficient Task Scheduling in Cloud Environment
Shamita Phutane1, Ankith Poojari2, Saurabh Nyati3, Bhavna Arora4
1,2,3BE student, Computer Department, Atharva College, Mumbai University
4 Professor, Computer Department, Atharva College, Mumbai University
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Cloud data centers consume large amounts of
cooling power. Previous research was primarily based on
the provision of mission plans, optimizing both
computational power and overall performance gains.
However, with the development of cloud facts, user needs
are changing and cannot be met by traditional scheduling
algorithms. One of the necessities is to maintain the value of
cooling the facts in the middle as much as possible.
According to literature reviews, the value of cooling power
is half a million dollars. Executing value in a cloud
environment requires a good mission planning approach.
Temperature affects not only reliability, but also the overall
performance, performance, and cost of embedded systems.
This section describes the heat-aware task assignments and
set of rules for cloud data centers. Mission plans are built in
ways that not only reduce computational costs but also
reduce cooling. Rigorous simulations were performed and
compared to state-of-the-art graphics algorithms.
Experimental results show that the rules of thermal task
planning are superior to other strategies. This project aims
to provide cloud data centers with a heat-aware task
scheduling approach and improve data center performance
by adopting task scheduling techniques. The basic
motivation is to provide a valuable and powerful green
algorithm through the fact center.
Key Words: Energy-performance, Green cloud, Multi
resource, Task scheduling, energy-aware systems.
1. INTRODUCTION
Loud computing provides users with convenient, on-
demand network access to shared computing resources,
making it a promising standard for service providers and
users. In most cases, these resources are heterogeneous,
inherently allocated, and can be acquired and rejected
immediately with nominal supervision or support from
the service provider. The increasing cost of use in data
centers and the associated environmental hazards have
increased the need for intensity-aware computing. Power
consumption has emerged as an important issue in cloud
computing, but some work has been done to manage
power consumption efficiently. It becomes more
important while other issues are raised in cloud
computing. The cloud computing environment manages a
large number of clusters that hold running services in the
data center. Services are provided in the form of Platform
as a Service (РааS), Software as a Service (SаAS), and
Infrastructure as a Service (IAAS). To take into account the
amount of power consumed by overloaded and
underutilized computer systems, you need a good
algorithm for task planning that can use resources
efficiently. The reason is that most servers remain idle or
resources are being used inefficiently. It needs а seriоus
соnсern fоr sосiety сlоud dаtа сenter lоt. To improve the
efficiency of energy consumption, we propose a
framework for deep reinforcement learning. This
document uses an energy-efficient depth forced learning-
based job scheduling framework. Reinforcement learning
is part of machine learning often used to solve automation
problems. The main feature of reinforcement learning is
that no pre-configured data is required for training. The
agent itself needs to investigate the training data. Forced
learning is used to maximize output. In our case, this is
energy optimization. For agents specific steps are required
to maximize the required functionality defined, this leads
to optimization of energy. This is the reason reinforcement
learning makes it an ideal learning system for energy-
efficient work planning on a cloud platform.
2. REVIEW OF LITERATURE
In a cloud computing environment, user services always
require heterogeneous resources (CPU, I / O, memory,
etc.). but also reduce energy consumption and the time it
takes to complete a user's order.Therefore, scheduling and
load balancing techniques are very important for an
efficient cloud setup with limited resources. Cloud
computing task scheduling has been tackled by many
researchers in the past.
Virtualization has become a powerful technology in data
centers and cloud computing [3]. This technology allows
you to share physical servers or hosts between different
virtual machines through a virtual machine monitor
(VMM) or hypervisor to increase resource utilization. One
of the main benefits of virtualization is Dynamic VM
Integration (DVMC), which is used in data centers to
reduce energy consumption. Power efficiency is achieved
by consolidating VMs into a minimal number of hosts and
hibernating idle hosts. DVMC uses the live migration
feature provided by virtualization technology to transfer
running VMs from one host to another. VM migration is
useful if your host commits significantly under or over.
Therefore, resource management policies become more
flexible through migration operations. This paper [4]
proposes a hybrid assignment scheduling set of rules
named FMPSO that is primarily based totally on the Fuzzy

device and the Modified Particle Swarm Optimization
method to decorate load balancing and cloud throughput.
FMPSO approach at the start considers 4 changed pace
updating strategies and roulette wheel choice method to
decorate the worldwide seek capability. Then, it makes
use of crossover and mutation operators to conquer a few
drawbacks of PSO consisting of nearby optima. Finally, this
schema applies a fuzzy inference device for health
calculations. The enter parameters for the proposed fuzzy
device are the period of tasks, the velocity of CPU, length
of RAM, and overall execution time. By including those
fuzzy systems, the FMPSO approach achieves the purpose
of minimizing the execution time and aiding usage. The
FMPSO set of rules, the usage of the CloudSim toolkit, and
simulation effects reveal that the proposed approach has a
higher overall performance in phrases of makespan,
development ratio, imbalance degree, efficiency, and
overall execution time evaluating different approaches.
The purpose of these paintings is to distribute Big
information for allotted processing withinside the Cloud in
a manner to lower the overall strength intake of the Cloud,
without scarifying the utility`s overall performance.
Therefore, they develop a strength-conscious assignment
scheduler (EATS) whose purpose is to lessen the
electricity intake of the applied resources and the
processing time of a utility. Divisible load packages are a
category of packages that may be divided into impartial
duties and assigned to distributed sources and not using
synchronization and inter-duties communication. The
divisible load takes place in lots of scientiﬁcand
engineering packages to investigate and method Big
information, inclusive of looking for a pattern,
compression, generic search packages, multimedia and
video processing, picture processing, and information
mining. The distribution of processing is supposed to
boost the overall performance of the allotted utility as
compared to its sequential execution. The advanced
scheduler to boost the overall performance of divisible
load packages in a Cloud Computing environment. The
scheduler follows a linear-version method and no longer
recollects the energy consumption of the cloud. In this
work, the expansion of non-linear programming is
primarily based scheduling version which aims at
optimizing the overall performance of the utility and the
strength intake of the underlying sources. This indicates
that idle, and underutilized sources devour a massive
quantity of strength as compared to that of complete use.
The version takes into attention this commentary to
optimize strength intake in the course of distributed
computing.
3. PROPOSED SYSTEM
Many cloud architectures are available in the market, but
most of them are focusing only on cloud architectural
performance. However, in further development of the
cloud data center, the needs of the user can not be
diversified and can not be achieved by traditional planning
algorithms. The Cloud Data Center is equipped with
hostile/cold floor tiles on the high floor. Соmputer room
соnditininer (сrас) is air conditioning for data centers and
cold air. Air flows through the front end, absorbs heat as it
flows through these frames, and exits in front of the rear
end of the frame. Warm air leaves the rack and is returned
to the air-conditioned air intake above the warm aisle.
Each rack is equipped with many cases, and each case is
equipped with many computing resources such as servers
and network equipment. A large amount of heat is
released, polluting the environment and leading to global
warming.
The main purpose of this post is to develop a cloud data
center heat-aware task scheduling strategy and apply task
scheduling techniques to improve data center
performance. According to [3], the task scheduling
algorithm finds suitable resources for the task sent to a
cloud data center according to the corresponding
characteristics. It is speculated that the proposed method
can minimize proper scheduling of data center tasks by
increasing thermal recirculation, helping the data center
operations in energy-efficient modes and achieve savings.
Will be done.
Our project uses a greedy approach to develop
temperature-based energy-efficient task scheduling
algorithms. The main purpose of this post is to develop a
cloud data center heat-aware task scheduling strategy and
adopt task scheduling techniques to improve data center
performance. According to [3], the task planning algorithm
finds suitable resources for the task sent to the cloud data
center according to the appropriate characteristics. In the
proposed method, the increased thermal recirculation can
minimize the proper planning of tasks in the data center,
and it is speculated that it will help the data center to
operate in an energy-efficient model to save. The proposed
model is based on DRL where the temperature is
considered along with the other parameters. The
temperate parameter will help to plan the execution of
tasks by the thermal efficiency of the server.
1. With the advancement of cloud data centers, the
needs of users are diversifying, which is achieved
by the JARS algorithm.
2. To make cloud data centers green and reduce
carbon emissions, energy consumption must also
be minimized.
3. Greening, the cloud architecture can also
minimize the waste generated from the cloud data
center
4. Task scheduling is designed to not only reduce
computational costs.

3.1. Model Description
Orders submitted by users are independent of each other.
There are different servers in the cloud, and each server
has only a VM. Also, each VM can only run one job at a
time. These VMs are used for online work and planning.
VM was assigned to a position, with no prior knowledge of
the profession
Many public cloud providers, such as Amazon Web
Services in the cloud and Microsoft Azure, have several
alternative options for VMs. If you are considering an AWS
EC2 instance. It consists of a faster CPU with more
processing power and a slower midrange CPU with
different processing power. For example, in the case of,
small jobs run smoothly with the following jobs, providing
the CPU VM with a more powerful CPU to run in more
intensive jobs. These instances with different prices on
public cloud platforms also differ because these are
measured on a pay-as-you-go basis.
The job scheduling process is somehow professionally
tracked and jobs are scheduled for specific VMs available
in the resource pool on which this energy-efficient model
is based. Multiple user input jobs in real-time. Let us
explain this further. The number of job queues in the
resource pool, for temporary positions. A specific runtime
for a job. VMs cannot consider different jobs because other
VMs only run one job. If you interrupt the process, you will
not be able to do the work in progress and you will not be
able to purchase it in advance.. In our model, we will be
having S servers and each server will be having exactly
one VM on it,
VMj = { VM_idj , VM_Comj , Ej } (1)
User jobs are defined by job arriving time, computational
power, and user request time i.e QoS.
Jobi = { Job_idi , arrival_Timei , Job_sizei , QoSi } (2)
When the job queue is empty and a job arrives, will be
executed immediately even if a job occurs there is a job
queue and there is a job with an arrival time before it all
jobs prior to the last job will only run once it is planned
soon:
Ti = Ti wait + Ti_exe (3)
The job when completes its execution on a VM and returns
to the end-user and the time taken by the job to get
completed can be evaluated as
Ti_leave = arrival_Timei + Ti (4)
4. Implementation
Looking back on previous research on this technical topic
arena One can realize some beliefs the study group holds
about reinforcement learning. The current strengthening
discipline is that it takes a long time to learn something
together. Some people think that reinforcement learning
techniques can only be used in practice. The method is
either pre-training or transfer in the simulator. Knowledge
learned between different domains when compared to
other machine learning methods, this philosophy limits
the technology used. Some people hesitate to answer real
questions and make choices for further research in the
field. This may be the reason this area has been very
successful in video games such as strategy games like
Atari or Go with known rules predefined. It contains all the
pre-processing required for other proposed algorithms.
First, create a network topology that defines the number
of data centers, hosts, and VMs. Then configure all the VMs
and the value of each VM according to its characteristics.
Third, determine the path of your network topology.
Lastly calculate the total number of resources (CPU,
memory, memory) in each path of the network total each
resource. The proposed task describes the planner as an
agent.
4.1. Task Scheduling algorithm
The following algorithm is used in our model to predict
the best-case scenario considering the energy
consumption using both deep Q learning techniques and
offline decision techniques such as experience replay and
target networks to increase the efficiency
Аlgоrithm 1: Job And Resources Scheduling Algorithm
(JARS)
for episode =1, do
Reset cloud server environment to initial
state
with probability e randomly choose
an action a j ;
Schedule job j according to action a j,
receive
reward r j, and observe state transition at
next
decision time t j + 1 with a new state s j +
1
Store transition j
if j > t and j ≡ 0 then
if j ≡ 0 then
Reset
end if
randomly select samples state
from states
for each transition in the state do
target k = rewards
end for
end if
end for

Algorithm 2: Temperature Seeker Algorithm
fоr episode in 1, dо
if check enough resources available on
the
server for a substitute task then
саlсulаtetemрerаture(Counter,
ti) ;
end if
if (temperаture less than Mtemp) then
Allocate tasks and resources;
Mtemp=temрerаture;
end if
end for
4.3. DRL-DQN Approach
At each time step t, the agent monitors the current state of
the environment and takes action. From this interaction,
the environment tells the agent how well the agent is
doing. The agent's goal is to maximize rewards over time.
This issue schedules tasks based on CPU and memory
allocation. The selection is made from a stack of tasks. The
schedule only considers the CPU and memory, does not
fragment the resources, and assumes that the resources
consumed by the task are known in advance
State: The state is defined as all tasks in the queue. In this
case, each stack has a total of s tasks. You can choose from
a total of s states.
Actions: For each action, the agent can select a task or
nothing. If the agent does not select anything, the selection
phase ends and all selected tasks are processed. . You
cannot reselect the selected task, so each time you make a
selection, the number of possible actions is reduced until
the end of the selection phase.
Reward: The environment works with agent actions. The
environment returns rewards and new conditions. The
new state is the entire task queued in the previous
selection minus the selected task. If the choice is wrong,
the environment will punish the agent by setting the
reward
Chart -1: DRL working
Negative value. The total number of resources is important
for the correct selection of rewards
Two different Neural networks i.e target network and
evaluation network are used in our model. The Target
Network is considered a temporarily frozen network. The
parameters of the target network are copied from the
evaluation network after every step. From there, a search
scheme (usually Epsilon Leady) is used to stochastically
choose between the best Q-factor action and random
action. At a higher level, deep learning works as follows: It
uses the current policy to collect samples and store them
in the replay buffer, randomly sample the experience from
the replay buffer (so-called experience replay), and use
the sample experience to update and repeat the Q
network. Successive experiences are strongly
(temporarily) correlated. Statistical learning and
optimization tasks require an independent distribution of
data. In other words, you don't want the data you're
feeding to be correlated. A random sample of experience
decomposes this temporal correlation of behavior and
distributes/averages it into many previous states. This
avoids large fluctuations and biases (problems that can
arise from correlation data) in the Model. To update the Q
network, you need to minimize the root mean square
error between the Q target value (according to the
Bellman equation) and the actual Q output. To reduce
errors, this means that the results of the current policy
will become more and more similar to the true Q value.
Therefore, use the loss function defined above to perform
the gradient step of the loss function.

Chart -2: System Architecture
5. Result and Conclusion
This figure shows the success rate in terms of energy
efficiency. We have defined various test series. One is a
small (1000) workstation, a medium (5000) workstation,
and a large (10000) workstation. All jobs of all sizes will
be online at different intervals. Each generated job does
not depend on any other job. In addition, each job has
different parameter values that are efficiently scheduled
with energy consumption as well as parameters such as
QoS. The comparison is based on the energy factor.
Compared to other baselines such as round-robin,
random, and earlies algorithms, JARS outperforms them
with 80% consistent accuracy for all three workloads,
Small, Medium, Large. Other baselines achieved less than
55% accuracy on any workload. The round-robin policy
schedules jobs instantly without making a decision. Job
allocation is very fast, so if you have a small number of
VMs, the response time will be very fast. Round-robin
indicates 51-currency when the workload is light and 48-
currency when the workload is increased. Medium
workload accuracy, 45 ° accuracy for heavy workloads.
The Random Policy, as the name implies, randomly assigns
jobs to different servers, makes no decisions, and
performs random actions, similar to Round Robin, so
response times are very fast. The average power efficiency
of the random policy was 47%, and the workload adopted
was the lightest of the various policies used in the model.
The original policy selected the first scheduled job on the
server with the highest power consumption and queued
other jobs, averaging across the workload, exceeding the
round-robin and random policies. You can see that the
policy is more efficient than random and round-robin, but
still very inaccurate. Deep Reinforcement Learning, on the
other hand, proves an average accuracy of over 80% for
different workloads and runs the best across the four
algorithms used. JARS response times are longer than the
other three, but scheduling jobs on a server-by-server
basis requires careful decision-making. JARS is superior to
the other three algorithms in that it uses an intelligent job
scheduling mechanism to minimize power consumption
and optimize power consumption. Reinforcement learning
requires a long way of improvement and development by
mitigating problems, which ultimately leads to great
research value on this subject. It will be a new field of
study. Reinforcement learning will be an implementable
technology that will bring great perspectives to dynamic
planning by strengthening the direction of many studies
where the work of data centers and users will change
dynamically shortly. In a changing environment, cloud
computing resources need to be optimally operated.
Therefore, the generic JARS-based algorithm is suitable for
cloud computing because it uses resources effectively and
reduces power and margins. Experimental results show
that the proposed method using the largest application
and a scheduling algorithm for random scheduling is
excellent for further research. The optimized model needs
to add other necessary goals such as cost, bandwidth, and
loss balance. We also need to focus more on robust
algorithms.
Chart -3: Final Output
REFERENCES
[1]Farahnakian, Fahimeh & Pahikkala, Tapio & Liljeberg,
Pasi & Plosila, Juha & Tenhunen, Hannu. (2015).
Utilization Prediction Aware VM Consolidation Approach
for Green Cloud Computing. 10.1109/CLOUD.2015.58.
[2]Mansouri, Name & Mohammad Hasani Zade, Behnam &
Javidi, M.M.. (2019). Hybrid Task Scheduling Strategy for
Cloud Computing by Modified Particle Swarm
Optimization and Fuzzy Theory. Computers & Industrial
Engineering. 130. 10.1016/j.cie.2019.03.006.
[3]Ismail, Leila & Fardoun, A.A.. (2016). EATS: Energy-
Aware Tasks Scheduling in Cloud Computing Systems.
Procedia Computer Science. 83. 870-877.
10.1016/j.procs.2016.04.178.

[4]Tang, Qinghui & Sandeep, Kornepati & Varsamopoulos,
Georgios. (2007). Thermal-Aware Task Scheduling for
Datacenters through Minimizing Heat Recirculation.
Proceedings - IEEE International Conference on Cluster
Computing, ICCC. 129-138. 129-138.
10.1109/CLUSTR.2007.4629225.
[5]Donald, James & Martonosi, Margaret. (2006).
Techniques for Multicore Thermal Management:
Classification and New Exploration. ACM Sigarch
Computer Architecture News. 34. 78-88.
10.1109/ISCA.2006.39.
[6]Chen, Y., Das, A., Qin, W., Sivasubramaniam, A., Wang, Q.,
& Gautam, N. (2005). Managing server energy and
operational costs in hosting centers. ACM SIGMETRICS
Performance Evaluation Review, 33(1), 303–314.
https://doi.org/10.1145/1071690.1064253
[7]Springer, R., Lowenthal, D.K., Rountree, B., & Freeh,
V.W. (2006). Minimizing execution time in MPI programs
on an energy-constrained, power-scalable cluster. PPoPP
'06.
[8]Bash, Cullen & Forman, George. (2007). Cool Job
Allocation: Measuring the Power Savings of Placing Jobs at
Cooling-Efficient Locations in the Data Center.. 363-368.
[9]Reinforcement Learning Techniques for Optimal Power
Control in Grid-Connected Microgrids: A Comprehensive
Review - Scientific Figure on ResearchGate. Available
from: https://www.researchgate.net/figure/A-data-flow-
diagram-for-a-batch-reinforcement-learning-algorithm-A-
memory-location-is_fig3_346110033 [accessed 30 Mar,
2022]
[10]T. Heath, B. Diniz, E. V. Carrera, W. Meira Jr, and R.
Bianchini, “Energy conservation in heterogeneous server
clusters,” in Proceedings of the tenth ACM SIGPLAN
symposium on Principles and practice of parallel
programming.ACM,2005,pp.186–195.
[11]P. Ranganathan, P. Leech, D. Irwin, and J. Chase,
“Ensemble-level power management for dense blade
servers,” in ACM SIGARCH Computer Architecture News,
vol. 34, no. 2. IEEE Computer Society, 2006, pp.66–77.
[12]Shu et al.: A novel energy-efficient resource allocation
algorithm based on immune clonal optimization for green
cloud computing. EURASIP Journal on Wireless
Communications and Networking 2014 2014:64
[13]P.Bohrer,D.Cohn,E.Elnozahy,T.Keller,M.Kistler,C.Lefur
gy, R. Rajamony, F. Rawson, and E. Hensbergen, “Energy
conservation for servers,” in Workshop on Power
Management for Real-Time and Embedded Systems, 2001.

Energy-Efficient Task Scheduling in Cloud Environment

More Related Content

Similar to Energy-Efficient Task Scheduling in Cloud Environment (20)

More from IRJET Journal (20)

Recently uploaded (20)

Energy-Efficient Task Scheduling in Cloud Environment