Date of Award

Summer 2012

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computational Analysis and Modeling

First Advisor

Chakzhad Leangsuksun

Abstract

With the enormous number of computing resources in HPC and Cloud systems, failures become a major concern. Therefore, failure behaviors such as reliability, failure rate, and mean time to failure need to be understood to manage such a large system efficiently.

This dissertation makes three major contributions in HPC and Cloud studies. First, a reliability model with correlated failures in a k-node system for HPC applications is studied. This model is extended to improve accuracy by accounting for failure correlation. Marshall-Olkin Multivariate Weibull distribution is improved by excess life, conditional Weibull, to better estimate system reliability. Also, the univariate method is proposed for estimating Marshall-Olkin Multivariate Weibull parameters of a system composed of a large number of nodes. Then, failure rate, and mean time to failure are derived. The model is validated by using log data from Blue Gene/L system at LLNL. Results show that when failures of nodes in the system have correlation, the system becomes less reliable.

Secondly, a reliability model of Cloud computing is proposed. The reliability model and mean time to failure and failure rate are estimated based on a system of k nodes and s virtual machines under four scenarios: 1) Hardware components fail independently, and software components fail independently; 2) software components fail independently, and hardware components are correlated in failure; 3) correlated software failure and independent hardware failure; and 4) dependent software and hardware failure. Results show that if the failure of the nodes and/or software in the system possesses a degree of dependency, the system becomes less reliable. Also, an increase in the number of computing components decreases the reliability of the system.

Finally, an economic model for a Cloud service provider is proposed. This economic model aims at maximizing profit based on the right pricing and rightsizing in the Cloud data center. Total cost is a key element in the model and it is analyzed by considering the Total Cost of Ownership (TCO) of the Cloud.

Share

COinS