An Adaptive Framework for the Execution of Data-Intensive MapReduce Applications in the Cloud

Abstract

Cloud computing technologies play an increasingly important role in realizing data-intensive applications by offering a virtualized compute and storage infrastructure that can scale on demand. A programming model that has gained a lot of interest in this context is MapReduce, which simplifies processing of large-scale distributed data volumes, usually on top of a distributed file system layer. In this paper we report on a self-configuring adaptive framework for developing and optimizing data-intensive scientific applications on top of Cloud and Grid computing technologies and the Hadoop framework. Our framework relies on a MAPE-K loop, known from autonomic computing, for optimizing the configuration of data-intensive applications at three abstraction layers: the application layer, the MapReduce layer, and the resource layer. By evaluating monitored resources, the framework configures the layers and allocates the resources on a per job basis. The evaluation of configurations relies on historic data and a utility function that ranks different configurations regarding to the arising costs. The optimization framework has been integrated in the Vienna Grid Environment (VGE), a service-oriented application development environment for providing applications on HPC systems, clusters and Clouds as services. An experimental evaluation of our framework has been undertaken with a data-analysis application from the field of molecular systems biology.