H-WorD: Supporting Job Scheduling in Hadoop with Workload-Driven Data Redistribution

摘要

Today's distributed data processing systems typically follow a query shipping approach and exploit data locality for reducing network traffic. In such systems the distribution of data over the cluster resources plays a significant role, and when skewed, it can harm the performance of executing applications. In this paper, we address the challenges of automatically adapting the distribution of data in a cluster to the workload imposed by the input applications. We propose a generic algorithm, named H-WorD, which, based on the estimated workload over resources, suggests alternative execution scenarios of tasks, and hence identifies required transfers of input data a priori, for timely bringing data close to the execution. We exemplify our algorithm in the context of MapRe-duce jobs in a Hadoop ecosystem. Finally, we evaluate our approach and demonstrate the performance gains of automatic data redistribution.