Identifier

etd-06282012-165206

Degree

Master of Science in Computer Science (MSCS)

Department

Computer Science

Document Type

Thesis

Abstract

The volume and complexity of data that must be analyzed in scientific applications is increasing exponentially. Often, this data is distributed; thus, the ability to analyze data by localizing it will yield limited returns. Therefore, an efficient processing of large distributed datasets is required, whilst ideally not introducing fundamentally new programming models or methods. For example, extending MapReduce - a proven effective programming model for processing large datasets, to work more effectively on distributed data and on different infrastructure (such as non-Hadoop, general-purpose clusters) is desirable. We posit that this can be achieved with an effective and efficient runtime environment and without refactoring MapReduce itself. MapReduce on distributed data requires effective distributed coordination of computation (map and reduce) and data, as well as distributed data management (in particular the transfer of intermediate data units). To address these requirements, we design and implement Pilot-MapReduce (PMR) - a flexible, infrastructure-independent runtime environment for MapReduce. PMR is based on Pilot abstractions for both compute (Pilot- Jobs) and data (Pilot-Data): it utilizes Pilot-Jobs to couple the map phase computation to the nearby source data, and Pilot-Data to move intermediate data using parallel data transfers to the reduce computation phase. We analyze the effectiveness of PMR over applications with different characteristics (e. g. different volumes of intermediate and output data). Our experimental evaluations show that the Pilot abstraction for data movement across multiple clusters is promising, and can lower the execution time span of the entire MapReduce execution. We also investigate the performance of PMR with distributed data using a Word Count and a genome sequencing application over different MapReduce configurations. We find that PMR is a viable tool to support distributed NGS analytics by comparing and contrasting the PMR approach to similar capabilities of Seqal and Crossbow, two Next Generation Sequencing(NGS) Hadoop MapReduce based applications. Our experiments show that PMR provides the desired flexibility in the deployment and configuration of MapReduce runs to address specific application characteristics and achieve an optimal performance, both locally and over wide-area multiple clusters.

Date

2012

Document Availability at the Time of Submission

Release the entire work immediately for access worldwide.

Committee Chair

Jha, Shantenu

Share

COinS