Metodología de predicción de rendimiento de aplicaciones MapReduce iterativas sobre una nube híbrida

Abstract

Cloud-bursting (complement on-premise virtual machines with temporary off-premise virtual machines) has seen a rapid adoption among big data analitycs users. However, in these type of deployments, performance is a challenge. On the one hand, Hadoop MapReduce is designed to schedule its tasks close to the data, on the other hand, there is a communication network between both clouds with a limited bandwidth. This thesis addresses these challenges through several contributions that converge towards a holistic solution. Specifically, this thesis provides an analysis tool that allows the subsequent proposal of strategies to efficiently use the locality of the data. Based on these strategies, which make feasible the use of iterative MapReduce applications, it contributes with a completion time prediction methodology, together with an economic cost model.