Volume 9, Issue 10, October– 2024 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24OCT092

Impact of Autoscaling on Application

Performance in Cloud Environments
Shankar Dheeraj Konidena

Abstract:- The current paper studies the impact of I. INTRODUCTION

Autoscaling on application performance in Cloud
computing environments. Cloud computing is one of the Autoscaling in cloud computing is adjusting the
most encouraging innovations due to its vast company resources horizontally by adding/removing VMs
applications. Predictive autoscaling is an advanced and modifying CPU/ memory usage in response to changes
technique that aims to address the challenges in the in workload. The topic of devising auto-scaling mechanisms
autoscaling trends for large-scale systems. To account consists of subtopics that address specific needs and
for the Quality of Service to the customer, features like techniques of auto-scaling, including virtualization,
load balancing according to the workload demand, comparison, workload monitoring, hybrid/multi-clouds,
understanding resource allocation and utilization, and bandwidth, integrated storage, network and computing, and
dynamic decision-making are vital to any cloud self-scaling frameworks. [2] Considering the Cloud is a
computing application. This paper interprets these dynamic and uncertain environment, Reinforcement
challenges and reviews a meta-reinforcement learning Learning (RL) serves as a good candidate for autoscaling
approach for predictive autoscaling in cloud since it is capable of learning transparent (with no human
environments. A novel RL-based predictive autoscaling intervention), dynamic (no static plans), and adaptable
approach on a popular large-scale digital payment (constantly updated) resource management policies to
platform system, Alipay, is compared with the existing execute scaling actions. [1]. To quickly adapt to new
models such as Autopilot and FIRM. The aim is to workloads by learning common patterns and reducing the
conduct a detailed analysis of performance metrics need to train separate models, meta-learning aids in
before and after autoscaling actions, aiming to identify generalizing across different applications. Various
optimal scaling strategies that minimize response time applications and workloads impact CPU utilization
and maximize resource utilization without over- differently, which, in turn, complicates scaling decisions.
provisioning. Meta models address the commonalities between workloads
and, at the same time, keep the specifics of the application
Keywords:- Predictive Autoscaling, Application unchanged, portraying flexibility. Dynamic workloads, CPU
Performance, Cloud Applications, Machine Learning, utilization, and efficient allocation of resources are prevalent
Reinforcement Learning. challenges in Autoscaling.

Fig 1 Architecture of Autoscaling

IJISRT24OCT092 www.ijisrt.com 1
Volume 9, Issue 10, October– 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24OCT092

II. BACKGROUND or workloads that vary predictably with daily or weekly

cycles. Without predictive autoscaling, an auto scaler can
 Quality of Service and Service Level Arguments: only scale a group reactively based on observed changes in
Large enterprise software systems such as eBay, load in real time. With predictive Autoscaling enabled, the
Priceline, Amazon, and Facebook must provide high auto scaler works with real-time and historical data to cover
assurance in Quality of Service (QoS) metrics such as the current and forecasted load. Forecasts are refreshed
response times, high throughput, and service availability to every few minutes (faster than competing clouds) and
their users. With such assurances, service providers of these consider daily and weekly seasonality, leading to more
applications stand to retain their user base and, hence, their accurate forecasts of load patterns. [blog]
revenues. Typically, customers maintain Service Level
Agreements (SLAs) with service providers for the QoS  Reinforcement Learning -Artificial Intelligence
property. Catering to the SLA while keeping costs low is Cloud computing has revolutionized the IT industry by
challenging for such enterprise systems due primarily to the offering Infrastructure as a Service (IaaS), Platform as a
varying number of incoming customers. [2]. Various Service (PaaS), and Software as a Service (SaaS), providing
possibilities and approaches could be designed to solve scalable computing, network, and storage capabilities
different problems occurring in cloud architecture, through a pay-per-use model. Machine Learning (ML)
especially in scaling arenas. Further sections delve deeper techniques emerged as critical components in optimizing
into the issues and necessitate suitable solution possibilities resource utilization and ensuring Quality of Service (QoS)
using Machine Learning and Neural Network techniques. in cloud environments. The evolution of cloud computing
has given rise to several paradigms tailored to specific
 Virtualization and Autoscaling application requirements, including edge computing, fog
Virtualization refers primarily to platform computing, mist computing, Internet of Things (IoT),
virtualization or the abstraction of physical resources for Software-Defined Networking (SDN), cybertwin, and
users. These physical resources are regarded as a pool of Industry 4.0. Delivering customer-centric services and
resources; thus, they can be allocated on demand. collectively enhancing the Quality of Experience (QoE) for
Computing at the scale of the Cloud system allows users to the end users are the paradigms for operating with cloud
access the enormous and elastic resources on demand. servers. ML techniques serve as a fundamental enabler for
However, user demand for resources can vary at different these emerging paradigms, addressing many challenges in
times, and maintaining sufficient resources to meet peak cloud computing. These include resource scheduling and
requirements can be a burden cost-wise. On the contrary, if provisioning, load balancing, Virtual Machine (VM)
the user maintains only minimal computing resources, the migration and mapping, task offloading, energy
resource is insufficient to handle the peak requirements. [8] optimization, workload prediction, and device monitoring.
Auto-scaling allows to scale your processed resources The integration of ML with cloud computing has
dynamically or as expected. Dynamic scaling: Starting and significantly improved the efficiency and effectiveness of
halting computing using various pre-characterized these operations.
conditions. In predictable scaling, one is sure about traffic
patterns, with the end goal being where traffic peaks each Despite the rapid advancements in this field, there
morning and goes down at some point of time. Auto-scaling remains a notable gap in the literature regarding
starts more Web servers toward the beginning of the day and comprehensive surveys that explore:
shuts down unnecessary ones in the prescribed timings amid
the cloud distribution. The load balancer feature in auto-  The integration of multi-paradigm architectures in cloud
scaling is a productive feature that enables handling the computing
overabundant load. Also, it is one of the viable techniques to  In-depth technical and analytical aspects of these
save costs and physical resources utilizing scaled-up and paradigms
scaled platforms dynamically according to the customers'  The pivotal role of ML techniques in emerging cloud
approaching traffic. Also, auto-scaling has dynamic edge computing paradigms.
strategies that predict required resources based on
anticipated values. The auto-scaling method gives on- This research aims to address this gap by thoroughly
demand assets according to workload in cloud computing investigating the integration of emerging cloud computing
circumstances (Lin et al.,2012; Ashraf et al., 2016). One of paradigms, focusing on the application of ML as a dominant
the critical characteristics of operating in the Cloud is problem-solving technology.
autoscaling, which elastically scales the resources
horizontally (the number of virtual machines (VMs)  Autopilot:
assigned is changed) or vertically (the CPU and memory A workload-based autoscaling method proposed by
reservations are adjusted) to match the changing workload. Google that builds the optimal resource configuration by
According to the timing of scaling, the autoscaling strategies seeking the best historical time window to match the current
can be divided into responsive and predictive strategies [1] window[1].
 Predictive Autoscaling
Customers use predictive autoscaling to improve
response times for applications with long initialization times

IJISRT24OCT092 www.ijisrt.com 2
Volume 9, Issue 10, October– 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24OCT092

 FIRM: with abnormal response time (RT) through SVM-based

An RL-based autoscaling method that solves the anomaly detection algorithms and adjusts multiple resources
problem through learning feedback adjustment in the online for the service through RL algorithms[1].
cloud environment. Specifically, FIRM finds applications

Fig 2 Taxonomy of Cloud Architecture

III. TECHNICAL RESULTS  Insight 1:

Deep time series models with notable success are
 Review of Insights: essential. Classical regression techniques or simple neural
The paper presents a novel RL-based predictive networks to forecast the workload are ineffective in
autoscaling approach. A deep, attentive periodic model for capturing periodicity or complex temporal dependencies.
multi-dimensional, multi-horizon workload prediction
provides high-precision and reliable workload information  Insight 2:
for scaling. The meta-learning model is used to train with Workload heterogeneity is discussed from two
the dynamic image to map the workload to CPU utilization, perceptions. (i) distinct applications have distinct mappings
with rapid adaptation to the changing environment, among workloads on CPU utilization; (ii) different workload
embedded to guide the learning of optimal scaling actions subtypes across an application have varied relationships
over thousands of online applications. The meta-model- with CPU utilization.
based RL algorithm enables safe and data-efficient learning.
Neural process and Markovian Decision process are  Insight 3:
significant aspects of the approach. There are three Insights A dynamic approach to decision-making is formed
involved in the process. when selecting the best resources, given a measure of CPU
utilization. The aim is to decide accurate resource allocation
(VMs) for the application according to the estimation of
CPU utilization.

Fig 3 Synergy Diagram of Workload Forecaster, CPU Utilisation Meta Predictor, and Scaling Decider.

IJISRT24OCT092 www.ijisrt.com 3
Volume 9, Issue 10, October– 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24OCT092

IV. RESULTS optimizing resource use, and showing strong decision-

making abilities. In terms of Scalability and RSME, the
 Auto Scaling via Meta Model-based RL model scaled across different numbers of applications (50,
The experimental results highlight the superior 100, 200, 500) while keeping prediction errors low: RMSE
performance of the proposed Meta Reinforcement Learning values increase slightly as the number of applications grows
approach for predictive autoscaling compared to traditional but remain relatively low (1.11 to 1.33), demonstrating
methods such as Informer, ConvTransformer, Autopilot, and accuracy, though applied to many applications. The
FIRM. In terms of Workload Prediction, the results approach scales well, handling large-scale cloud
prominently outperformed both Informer and environments without any drop in performance and
ConvTransformer, achieving an improvement of 25% in maintaining accuracy as the number of applications
accuracy with an MAE (Mean Absolute Error) of 1.10 and increased from 50 to 500. The model improved CPU
RSME (Root Mean Square Error) OF 112.9. This indicates stability by 20% and reduced cloud resource usage by 50%,
that the predictive model is highly accurate in forecasting exhibiting its effectiveness in real-world applications.
future workload spikes and dips, which is critical for
efficient autoscaling. Enhancement in the performance  Comparison of Autopilot, FIRM, and Proposed RL
attributes to capture complex temporal patterns. CPU Model:
stability rates are up to 0.95, and a 21% reduction in VM The following is the comparison table to differentiate
usage compared to Autopilot is seen in the scaling between Autopilot, FIRM, and Proposed RL Model and the
performance, achieving stable resource allocation. Accurate significance of the approach.
CPU utilization predictions enable better scaling decisions,

Table 1 Comparison of Autopilot, FIRM, and Proposed RL Model

Metric Autopilot FIRM Proposed RL Model Significance
Workload Prediction Higher (around 1.75) Moderate Lower (around 1.10) More accurate prediction leads
Accuracy (MAE) (around 1.50) to better resource planning.
CPU Utilization Higher (around 2.48) Moderate Lower (around 1.11) Better prediction of CPU
Prediction Accuracy (around 1.93) utilization helps in scaling
(RMSE) decisions.
CPU Utilization Fluctuates Moderate Highly stable (around More stable CPU utilization
Stability significantly stability 0.95) improves performance and
avoids resource waste.
VM Allocation Less efficient Moderate High efficiency Efficient resource use reduces
Efficiency (overshooting or efficiency (proactive scaling) costs.
undershooting VMs)
Failed Requests Higher (especially Fewer failed Fewest failed requests Better handling of peak loads
during spikes) requests ensures system reliability.
Scalability (RMSE for Higher RMSE (around Moderate Lower (around 1.11) The model scales better across
500 apps) 1.33) (around 1.25) many applications, showing
adaptability to large-scale
environments.
Cost Efficiency Higher costs due to Moderate cost Significant cost More resources are needed,
over-provisioning savings savings translating to lower operating
costs.
Real-World Deployed in Deployed in Deployed at scale Real-world validation shows
Performance (Alipay) controlled test (Alipay) practical benefits and
environments environments scalability.

V. CONCLUSION automatically select the best scaling strategy based on real-

time application behaviors and workload characteristics.
In conclusion, we can derive from the analysis that
carefully chosen autoscaling strategies can significantly REFERENCES
improve application performance. The RL-based meta-
learning approach is effective in CPU utilization in real- [1]. Xue, S., Qu, C., Shi, X., Liao , C., Zhu, S., Tan, X.,
world applications and scaling the digital platform, with a Ma, L., Wang, S., Hu, Y., Lei, L., Zheng, Y., Li, J.,
50% resource savings compared to the rule-based methods. & Zhang, J. (2022). A Meta Reinforcement Learning
RL techniques could be used in autoscaling to support Approach for Predictive Autoscaling in the Cloud. In
customers according to their Service Level Agreements, Proceedings of the 28th ACM SIGKDD Conference
assuring Quality of Service in a cost-friendly manner. Future on Knowledge Discovery and Data Mining (KDD
work should develop adaptive predictive models that ’22). https://doi.org/10.1145/3534678.3539063

IJISRT24OCT092 www.ijisrt.com 2
Volume 9, Issue 10, October– 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24OCT092

[2]. Roy, Nilabja & Dubey, Abhishek & Gokhale,

Aniruddha. (2011). Efficient Autoscaling in the
Cloud Using Predictive Models for Workload
Forecasting. Proceedings - 2011 IEEE 4th
International Conference on Cloud Computing,
CLOUD 2011. 500-507. 10.1109/CLOUD.2011.42.
[3]. Alipour, Hanieh & Hamou-Lhadj, Abdelwahab &
Liu, Yan. (2014). Analysing Auto-scaling Issues in
Cloud Environments.
[4]. Shahin, A. A. (2017). Automatic Cloud Resource
Scaling Algorithm based on Long Short-Term
Memory Recurrent Neural Network. ArXiv.
https://doi.org/10.14569/IJACSA.2016.071236
[5]. Amir Fazli, Amin Sayedi, Jeffrey D. Shulman (2018)
The Effects of Autoscaling in Cloud Computing.
Management Science 64(11):5149-5163.
[6]. J. H. Novak, S. K. Kasera and R. Stutsman, "Cloud
Functions for Fast and Robust Resource Auto-
Scaling," 2019 11th International Conference on
Communication Systems & Networks
(COMSNETS), Bengaluru, India, 2019, pp. 133-140,
doi: 10.1109/COMSNETS.2019.8711058. keywords:
{Delays; Cloud computing; Runtime; Current
measurement; Load modeling; Servers},
[7]. Arvindhan, M and Anand, Abhineet, Scheming a
Proficient Auto Scaling Technique for Minimizing
Response Time in Load Balancing on Amazon AWS
Cloud (March 15, 2019). International Conference on
Advances in Engineering Science Management &
Technology (ICAESMT) - 2019, Uttaranchal
University, Dehradun, India, Available at SSRN:
https://ssrn.com/abstract=3390801 or
http://dx.doi.org/10.2139/ssrn.3390801
[8]. Hung, Che-Lun & Hu, Yu-Chen & Li, Kuan-Ching.
(2012). Auto-Scaling Model for Cloud Computing
System. International Journal of Hybrid Information
Technology. 5. 181-186.
[9]. Jacob, S., & G. (2021, October 5). 10 ways Google
Cloud IaaS stands out. Google Cloud.
https://cloud.google.com/blog/products/compute/goo
gle-clouds-iaas-platform-is-a-powerful-choice

IJISRT24OCT092 www.ijisrt.com 5