Dynamic, adaptive, and optimal resource allocation and migration in cloud data centers

Date

2025-7-31

Author

Doğan, Onat Deniz

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

816
views

0
downloads

Cite This

Cloud Data Centers (CDCs) provide Infrastructure as a Service (IaaS) by allocating the computing resources of the Physical Machines (PMs) to Virtual Machines (VMs) according to the resource demands of the user. These demands are specified in the form of a service-level-agreement (SLA) and constitute the workload for the IaaS. The cloud service provider must meet the SLAs while efficiently allocating the resources and minimizing the power consumption. To this end oversubscription of the VMs is employed and resulting SLA violations are resolved by VM migration when necessary. Integer Linear Programming (ILP) has proven very effective in producing globally optimal placement and migration strategies under hard constraints. However, traditional ILP-based methods rely on statically defined parameters, which are unable to adapt to the temporal cloud workload variations. This thesis presents I2CLOUDMAN, a hybrid resource management approach combining the optimality of an ILP model with the real-time responsiveness of a Deep Q-Network (DQN). The DQN observes the up-to-date state of the data center and chooses the parameters of the ILP model dynamically, adjusting them to the workload conditions in each decision epoch. We implement a discrete-event simulation framework capable of replaying real traces for comparatively evaluating the performance of I2CLOUDMAN. To this end, we compare I2CLOUDMAN with three baselines: the static-parameter ILP model CLOUDMAN++, the PAPSO meta-heuristic, and Nova Scheduler of OpenStack using the Microsoft Azure 2017 VM workload traces. The outcome is that I2CLOUDMAN reduces the total power consumption by 9% compared to CLOUDMAN++ and 40% compared to PAPSO, while also achieving much fewer SLA violations than OpenStack's Nova Scheduler. These findings validate that learning-driven ILP tuning has the potential to realize scalable, energy-efficient, and SLA-aware resource management in modern cloud data centers.

Subject Keywords

Cloud data center resource management, Virtual machine placement and migration, ILP optimization, Deep reinforcement learning, Energy efficiency, SLA compliance

URI

https://hdl.handle.net/11511/115486

Collections

Graduate School of Natural and Applied Sciences, Thesis