Is it Time to Revisit Erasure Coding in Data-Intensive Clusters?

Data-intensive clusters are heavily relying on distributed storage systems to accommodate the unprecedented growth of data. Hadoop distributed file system (HDFS) is the primary storage for data analytic frameworks such as Spark and Hadoop. Traditionally, HDFS operates under replication to ensure data availability and to allow locality-aware task execution of data-intensive applications. Recently, erasure coding (EC) is emerging as an alternative method to replication in storage systems due to the continuous reduction in its computation overhead. In this work, we conduct an extensive experimental study to understand the performance of data-intensive applications under replication and EC. We use representative benchmarks on the Grid'5000 testbed to evaluate how analytic workloads, data persistency, failures, the back-end storage devices, and the network configuration impact their performances. Our study sheds the light not only on the potential benefits of erasure coding in data-intensive clusters but also on the aspects that may help to realize it effectively.