Resource Procurement in Cloud Systems
Open Access
- Author:
- Kamrava, Sepideh
- Graduate Program:
- Computer Science
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- August 26, 2016
- Committee Members:
- Bhuvan Urgaonkar, Thesis Advisor/Co-Advisor
George Kesidis, Committee Member
Bhuvan Urgaonkar, Committee Member - Keywords:
- Cloud Computing
Optimization
Resource Procurement
Reserved instances - Abstract:
- With the growth in complexity of problems needed to be solved, the old fashion systems cannot meet the user’s requirements. This niche has been identified by cloud providers offering enough resources with reasonable prices to handle any requests. The number of companies providing cloud services have been increasing in recent years. In order to compete with their rivals and attract more users, they provide more flexible options and plans. This dissertation focuses on one of the biggest cloud resource providers, i.e., Amazon's EC2 and investigate its structure and introduce different resource types and plans in details. The cloud users try to make an optimum decision on how they serve their demand with the available options. This decision making can be really complex given the degrees of freedom that exist in selecting among providers and their plans. Narrowing our attention on a smaller set of plans on Amazon's EC2 provider known as on-demand instances, reserved instances and the reserved instance marketplace, our goal is to derive a few guidelines on how different methods of assigning demands to the available plans can affect the users'total cost. To decrease the service cost, some optimization techniques including integer linear programming and dynamic programming are utilized to derive an optimal or sub-optimal decision for many general workloads. The results illustrate that how adding more plans to the system can further reduce the cost when the demand is noisier while for pure periodic workloads, the optimal cost can be achieved by only using on-demand and reserved instances. Moreover, a threshold based decision method is proposed to determine when a user should sell their available reserved instances to achieve the maximum benefit. A heuristic method is further used to incorporate the migration cost as an instance of hidden costs in the formulation. These techniques are applied to some real and synthetic workloads to confirm the expectations and derive useful guidelines.