Intelligent applications in several areas increasingly rely on big data solutions to improve their efficiency, but the processing and management of big data incur high costs. Although cloud-computing-based big data management and processing offer a promising solution to provide scalable and abundant resources, the current cloud-based big data management platforms do not properly address the high latency, privacy, and bandwidth consumption challenges that arise when sending large volumes of user data to the cloud. Computing in the edge and fog layers is quickly emerging as an extension of cloud computing used to reduce latency and bandwidth consumption, resulting in some of the processing tasks being performed in edge/fog-layer devices. Although these devices are resource-constrained, recent increases in resource capacity provide the potential for collaborative big data processing. We investigated the deployment of data processing platforms based on three different computing paradigms, namely batch processing, stream processing, and function processing, by aggregating the processing power from a diverse set of nodes in the local area. Herein, we demonstrate the efficacy and viability of edge-/fog-layer big data processing across a variety of real-world applications and in comparison to the cloud-native approach in terms of performance.

Cyberbullying is one of the harmful activities on social networks that particularly affects the mental well-being of adolescents. Recent research has focused on session-based approaches to cyberbullying detection, which consider various components of a social media session, including posts, comments, and user information. This holistic perspective deepens the understanding of cyberbullying. However, the effective modeling of interactions within a session by incorporating session components remains a significant challenge. Previous methodologies have often failed to capture the complex dependencies between session components, thereby limiting their efficacy in modeling session interactions within a graph structure. Consequently, critical patterns and insights within sessions may be overlooked. To address this problem, we initially modeled social media sessions and their interactions as heterogeneous graphs. Subsequently, we introduced a heterogeneous graph neural network framework for session-based cyberbullying detection. Our approach leverages a Heterogeneous Graph Neural Network to capture complex high-order semantic relationships from a learned heterogeneous graph. The model comprises three components: a feature encoder, a heterogeneous graph encoder, and a classifier that collaboratively classifies social media sessions as bullying or not. We conducted experiments using two real-world session-based datasets to assess the effectiveness of the proposed method. The results demonstrate remarkable performance on the Vine dataset, achieving an F1-score of 99.34. In addition, our method provided competitive results for the Instagram dataset. Furthermore, we carried out extensive experiments to underscore the significance of the graph components in the proposed approach.

Hadoop Distributed File System(HDFS) often experiences skew in data storage over time, mainly because of random data block allocation policy, datanode failure, replica reconstruction, and client activity, leading to utilization and load imbalance in the system. Although HDFS provides tools to rebalance the data in the cluster, balancer only considers balancing disk space utilization among nodes which re-allocates the data from highly utilized nodes to low utilized nodes. Thus, data access skew which is caused by piling a large amount of popular data in one node is not addressed in the default HDFS balancer. To address this issue, we present popularity-aware balancer based on node popularity score which spreads the popular data uniformly among datanodes, resulting in the balance of future access load balancing and reduction of hot spots in the cloud storage system. Simulation results demonstrate the promising benefits of proposed popularity-aware balancer by evaluating the uniform distribution of popular data across nodes without compromising the amount of data transfers and variance in disk space.

A long-term trend in high performance computing is the increasing number of nodes in parallel computing platforms, which entails a higher failure probability. Hence, fault tolerance becomes a key property for parallel application running on parallel computing systems. The message passing interface (MPI) is currently the programming paradigm and communication library most commonly used on parallel computing platforms. MPI applications may be stopped at any time during their execution due to an unpredictable failure. In order to avoid complete restarts of an MPI application because of only one failure, a fault tolerant MPI implementation is essential. In this paper, we propose a fault tolerant approach in cluster computing system. Our approach is based on reassignment of tasks to the remaining system and message logging is used for message losses. This system consists of two main parts, failure diagnosis and failure recovery. Failure diagnosis is the detection of a failure and failure recovery is the action needed to take over the workload of a failed component. This fault tolerant approach is implemented as an extension of the message passing interface.