Vers une plate-forme MapReduce tolérant les fautes byzantines

Byzantine faults are inherent in massive parallel computation, including those based on the MapReduce model. Yet, the current MapReduce framework implementations do not tolerate Byzantine failures. Therefore, it is not possible to verify if the final results of a MapReduce application are correct. We present in this article a MapReduce architecture where tasks are replicated aiming at ensuring the correctness of task execution results and isolation of faulty tasks. A preliminary performance study has evaluated some of our proposed replication mechanisms while a second one, conducted on top of a prototype that implements our architecture, has validated some of our choices, showing that it is possible to minimize the cost of Byzantine fault tolerance. MOTS-CLÉS : MapReduce, fautes byzantines, Hadoop, HDFS