Simulasi Autoscaling Untuk Optimalisasi Skalabilitas Dan Efisiensi Startup Dengan Microk8s Dan Reinforcement Learning Adaptif

Rohmat, Rohmat (2025) Simulasi Autoscaling Untuk Optimalisasi Skalabilitas Dan Efisiensi Startup Dengan Microk8s Dan Reinforcement Learning Adaptif. Masters thesis, Institute Teknologi Sepuluh Nopember.

Text 6047232001-Master_Thesis.pdf Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Aplikasi cloud-native pada tahap awal pengembangan rentan mengalami lonjakan trafik yang cepat dan tidak terduga, sementara keterbatasan sumber daya dan anggaran operasional tetap menjadi kendala utama. Kondisi ini menuntut mekanisme autoscaling yang tidak hanya efisien secara biaya, tetapi juga mampu menjaga kualitas layanan (Quality of Service/QoS) dan kepatuhan terhadap Service Level Agreement (SLA). Horizontal Pod Autoscaler (HPA) bawaan Kubernetes menggunakan pendekatan reaktif berbasis ambang batas metrik, yang sering kali kurang optimal dalam menghadapi perubahan beban yang dinamis dan abrupt. Penelitian ini mengusulkan pendekatan autoscaling adaptif berbasis Reinforcement Learning (RL) dengan arsitektur hibrida Deep Q-Network–Proximal Policy Optimization (DQN dan PPO). Fungsi reward dirancang untuk menyeimbangkan utilisasi CPU dan memori, penalti pelanggaran SLA dengan ambang latensi 200 ms, serta biaya operasional sebesar $0,10 per pod per step. Evaluasi dilakukan pada lingkungan MicroK8s berspesifikasi rendah dengan rentang skala 1–10 pod, menggunakan lima skenario beban lalu lintas sintetis independen (steady, gradual ramp-up, sudden spike, daily pattern, dan idle periods) yang dihasilkan menggunakan k6. Metrik kinerja utama dikumpulkan melalui Prometheus dan divisualisasikan menggunakan Grafana, kemudian dianalisis secara statistik. Hasil simulasi menunjukkan bahwa pendekatan Hybrid DQN-PPO secara konsisten memberikan kualitas layanan yang lebih baik dibandingkan HPA. Rata-rata waktu respons menurun sebesar 4,83%, dengan peningkatan signifikan pada metrik P95 response time hingga 19,61%. Selain itu, tingkat pelanggaran SLA berkurang sebesar 60,58%, dan perbedaan tersebut terbukti signifikan secara statistik (p < 0,001) dengan ukuran efek yang besar. Temuan ini mengindikasikan bahwa agen RL hibrida lebih mampu menjaga stabilitas performa layanan, khususnya pada kondisi beban tinggi dan lonjakan trafik. Secara keseluruhan, penelitian ini menegaskan bahwa Hybrid DQN-PPO merupakan alternatif autoscaling yang efektif bagi sistem cloud-native yang memprioritaskan kepatuhan SLA dan stabilitas layanan.
=====================================================================================================================================
Early-stage cloud-native applications are highly susceptible to sudden and unpredictable traffic surges, while operating under constrained computational resources and limited operational budgets. These conditions require autoscaling mechanisms that are not only cost-efficient but also capable of maintaining consistent service quality and strict Service Level Agreement (SLA) compliance. The Kubernetes Horizontal Pod Autoscaler (HPA) relies on reactive, threshold-based policies, which often struggle to adapt effectively to rapid and dynamic workload changes. This study proposes an adaptive autoscaling approach based on Reinforcement Learning (RL) using a hybrid Deep Q-Network–Proximal Policy Optimization (DQN–PPO) architecture. The reward function is designed to balance CPU and memory utilization, penalize SLA violations with a latency threshold of 200 ms, and account for operational costs at $0.10 per pod per simulation step. The approach is evaluated in a resource-constrained MicroK8s environment with a scaling range of 1–10 pods, across five independent synthetic traffic scenarios (steady, gradual ramp-up, sudden spike, daily pattern, and idle periods) generated using k6. Key performance metrics including response time, SLA violation rate, resource utilization, and cumulative cost are collected via Prometheus, visualized with Grafana, and analyzed using inferential statistical methods. Simulation results demonstrate that the Hybrid DQN–PPO approach consistently delivers superior service quality compared to HPA. Average response time is reduced by 4.83%, while P95 response time shows a substantial improvement of 19.61%. Furthermore, SLA violations decrease by 60.58%, with the differences proven to be statistically significant (p < 0.001) and exhibiting a large effect size. These findings indicate that the hybrid RL agent is more effective in maintaining performance stability, particularly under high-load conditions and abrupt traffic surges. Overall, the results highlight Hybrid DQN–PPO as a promising autoscaling alternative for cloud-native systems that prioritize SLA compliance and service stability. .

Item Type:	Thesis (Masters)
Additional Information:	https://github.com/rohmatmret/microk8s-autoscaling
Uncontrolled Keywords:	Autoscaling, Deep Q-Network , MicroK8s, Proximal Policy Optimization, Reinforcement Learning,
Subjects:	T Technology > T Technology (General) > T58.6 Management information systems
Divisions:	Interdisciplinary School of Management and Technology (SIMT) > 78201-System And Technology Innovation
Depositing User:	Rohmat Rohmat
Date Deposited:	21 Jan 2026 08:40
Last Modified:	21 Jan 2026 08:40
URI:	http://repository.its.ac.id/id/eprint/129980

Actions (login required)

View Item