Dissertations, Theses, and Capstone Projects

Date of Degree

5-2018

Document Type

Dissertation

Degree Name

Ph.D.

Program

Computer Science

Advisor

Kaliappa Ravindran

Committee Members

Jie Wei

Umit Uyar

Masum Hasan

Subject Categories

Computer Engineering

Keywords

Quality of Service (QoS), Service-level Management, Model-based Engineering, Machine Intelligence Tools, Assessment-as-a-service, Cloud-induced Uncertainties, Complex Systems Theory, Distributed/Replicated Services

Abstract

The issue of less-than-100% reliability and trust-worthiness of third-party controlled cloud components (e.g., IaaS and SaaS components from different vendors) may lead to laxity in the QoS guarantees offered by a service-support system S to various applications. An example of S is a replicated data service to handle customer queries with fault-tolerance and performance goals. QoS laxity (i.e., SLA violations) may be inadvertent: say, due to the inability of system designers to model the impact of sub-system behaviors onto a deliverable QoS. Sometimes, QoS laxity may even be intentional: say, to reap revenue-oriented benefits by cheating on resource allocations and/or excessive statistical-sharing of system resources (e.g., VM cycles, number of servers). Our goal is to assess how well the internal mechanisms of S are geared to offer a required level of service to the applications. We use computational models of S to determine the optimal feasible resource schedules and verify how close is the actual system behavior to a model-computed 'gold-standard'. Our QoS assessment methods allow comparing different service vendors (possibly with different business policies) in terms of canonical properties: such as elasticity, linearity, isolation, and fairness (analogical to a comparative rating of restaurants). Case studies of cloud-based distributed applications are described to illustrate our QoS assessment methods.

Specific systems studied in the thesis are: i) replicated data services where the servers may be hosted on multiple data-centers for fault-tolerance and performance reasons; and ii) content delivery networks to geographically distributed clients where the content data caches may reside on different data-centers. The methods studied in the thesis are useful in various contexts of QoS management and self-configurations in large-scale cloud-based distributed systems that are inherently complex due to size, diversity, and environment dynamicity.

Share

COinS