Satisfying strong application requirements in data-intensive cloud computing environments

Cho, Brian

Permalink

https://hdl.handle.net/2142/42371 Copy

Description

Title

Satisfying strong application requirements in data-intensive cloud computing environments

Author(s)

Cho, Brian

Issue Date

2013-02-03T19:36:39Z

Director of Research (if dissertation) or Advisor (if thesis)

Gupta, Indranil

Doctoral Committee Chair(s)

Gupta, Indranil

Committee Member(s)

• Abdelzaher, Tarek F.
• Godfrey, Philip B.
• Aguilera, Marcos K.

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Cloud Computing
• strong requirements
• bulk data transfer
• data consistency
• priority
• Hadoop

Abstract

In today's data-intensive cloud systems, there is a tension between resource limitations and strict requirements. In an effort to scale up in the cloud, many systems today have unfortunately forced users to relax their requirements. However, users still have to deal with constraints, such as strict time deadlines or limited dollar budgets. Several applications critically rely on strongly consistent access to data hosted in clouds. Jobs that are time-critical must receive priority when they are submitted to shared cloud computing resources. This thesis presents systems that satisfy strong application requirements, such as consistency, dollar budgets, and real-time deadlines, for data-intensive cloud computing environments, in spite of resource limitations, such as bandwidth, congestion, and resource costs, while optimizing system metrics, such as throughput and latency. Our systems cover a wide range of environments, each with their own strict requirements. Pandora gives cloud users with deadline or budget constraints the optimal solution for transferring bulk data within these constraints. Vivace provides applications with a strongly consistent storage service that performs well when replicated across geo-distributed data centers. Natjam ensures that time-critical Hadoop jobs immediately receive cluster resources even when less important jobs are already running. For each of these systems, we designed new algorithms and techniques aimed at making the most of the limited resources available. We implemented the systems and evaluated their performance under deployment using real-world data and execution traces.

Graduation Semester

2012-12

Permalink

http://hdl.handle.net/2142/42371

Copyright and License Information

Copyright 2012 Brian Cho

Owning Collections