[an error occurred while processing this directive]

Skip to Content

Computer Science Colloquia

Tuesday, November 22, 2011
Ming Mao
Advisor: Marty Humphrey
Attending Faculty: Sudhanva Gurumurthi (Chair), Jack Davidson, Jason Lawrence, and Teresa Culver

Olsson Hall, Room 228E, 3:00 PM

Ph.D. Defense Presentation
Cloud Auto-Scaling with Deadline and Budget Constraints

The cloud has become a fast-growing computing platform, which attracts more and more users. It offers on-demand computing power and storage capacity which can scale up or down based on the application performance requirements and business volume. Users only need to pay for what they use. The economies of scale behind cloud data centers can lower IT costs compared to on-premise hardware, infrastructure and management, especially for small and mid-size organizations and departments. The agility and cost advantages enable users to build highly scalable applications with limited budget. However, the cloud, as a black-box computing environment, does not offer any application-specific service level agreement, such as deadline, maximum response latency, minimum number of concurrent users supported, etc. Cloud developers are often not sure about the amount and the time to provision resources to achieve the target performance goals of their application. This is especially true when the application workload changes over time and experiences unexpected surges. The lack of efficient performance-aware cloud scaling mechanisms has become the main barrier that prevents users from making smart spending decisions aligned with their business needs and performance requirements in the cloud. The goal of this work is to build a performance- and workload-aware auto-scaling mechanism which can enable users to achieve their performance goals cost-efficiently in the cloud. The two main Quality of Service (QoS) metrics this work focuses on are deadline and cost. Deadline is chosen as the performance metric because it often reflects users preference on a jobs finish time, a direct indicator of system behaviors. Cost, on the other hand, reflects the financial capital users have to invest for a certain level of performance. Therefore, achieving user desired performance goals while minimizing the cost is the problem this work tries to solve.