Adoop: MapReduce for ad-hoc cloud computing
Mohammad Hamdaqa, Mohamed M. Sabri, Akshay Kr. Singh, Ladan Tahvildari · PolyPublie (École Polytechnique de Montréal) · 2015
MapReduce is a widely adopted distributed programming model for big data analytics. It facilitates processing large data sets using a dedicated cloud containing many worker nodes. For most implementations of MapReduce to work, only a few nodes may fail at any given time. This poses a challenge for organizations trying to harness the power of underutilized computing resources by provisioning cloud services on top of existing IT infrastructure. Exploiting ad-hoc clouds for MapReduce operations could yield significant advantages. It could reduce operational costs, improve resource utilization, and enable big data analytics. To the best of our knowledge, only few previous researchers have tried to optimize MapReduce for such volatile, non-dedicated, environments. This paper investigates how Hadoop --the most widely used open-source implementation of MapReduce-- can be optimized to run efficiently in ad-hoc cloud environments, despite the challenges these environments impose. To address these challenges, we present Adoop: a history-based scheduling approach to MapReduce, where the availability history of each node affects Hadoop scheduling decisions. Adoop maintains an availability and utilization based score of all the participating nodes, and dynamically re-adapts task assignments accordingly. A proof-of-concept implementation of Adoop has been provided and made publically available. Our initial experiments show that Adoop outperforms Hadoop in a simulated ad-hoc environment.