Performance modeling and resource management for MapReduce applications.

摘要

Big Data analytics is increasingly performed using the MapReduce paradigm and its open-source implementation Hadoop as a platform choice. Many applications associated with live business intelligence are written as complex data analysis programs defined by directed acyclic graphs of MapReduce jobs. An increasing number of these applications have additional requirements for completion time guarantees. The advent of cloud computing brings a competitive alternative solution for data analytic problems while it also introduces new challenges in provisioning clusters that provide best cost-performance trade-offs.;In this dissertation, we aim to develop a performance evaluation framework that enables automatic resource management for MapReduce applications in achieving different optimization goals. It consists of the following components: (1) a performance modeling framework that estimates the completion time of a given MapReduce application when executed on a Hadoop cluster according to its input data sets, the job settings and the amount of allocated resources for processing it; (2) a resource allocation strategy for deadline-driven MapReduce applications that automatically tailors and controls the resource allocation on a shared Hadoop cluster to different applications to achieve their (soft) deadlines; (3) a simulator-based solution to the resource provision problem in public cloud environment that guides the users to determine the types and amount of resources that should lease from the service provider for achieving different goals; (4) an optimization strategy to automatically determine the optimal job settings within a MapReduce application for efficient execution and resource usage. We validate the accuracy, efficiency, and performance benefits of the proposed framework using a set of realistic MapReduce applications on both private cluster and public cloud environment.