Faulty Node Detection in HDFS Using Machine Learning Techniques

The design of Hadoop has ability to elimination of fault tolerance, which consists of rescheduling the task on the defective nodes to run on other devices in the system. However, this strategy is ineffective if an error arises after most of the task has been completed. As a result, it is essential to make an early detection of the problem at the node to ensure that the resumption of the work will not result in a significant loss of both time and productivity. The ability to predict these problems provides us with the required time to move the workload onto different nodes, which helps to avoid data loss or processing time. In this paper, we propose an identification of faulty nodes from a large Hadoop distributed environment using machine learning techniques. Initially, we deployed one controller node and numerous data nodes as virtual machines in a distributed manner. The execution performs when the end-user submits a specific job to the controller node. The master node is the middleware controller that continuously communicates with other data nodes and assigns a task to each data node accordingly. However, this conventional process of HDFS that can generate data leakage or high computation whenever the specific node is heated or straggler. In our approach, we initially collect the log history of each data node and apply some statistical and a few machine learning algorithms to identify nodes' status. According to the achieved outcome of each node, we can decide to eliminate the specific node for task execution. We applied five machine learning algorithms in the extensive experimental analysis, including a Support Vector Machine (SVM). The SVM obtains 96.7% higher accuracy over the conventional machine learning classifiers for the entire execution.