The technological revolution, the huge sharing of data via social networks, web and mobile applications and IoT devices are generating a huge volume of data every day, commonly referred to as “Big Data”. To cope with Big Data and the challenges associated with their specific features, the last decade, a multitude of technologies and platforms have emerged to harness their potential. The community is still seeking a comprehensive and up-to-date comparative study of these tools. Such an experimentally-derived study is essential for enabling informed decision-making, fostering innovation, and ensuring that organizations can make the best choices when implementing Big Data solutions. In this paper, a multi-purpose experimental study was conducted. The primary objective is to provide an overview of today’s most popular Big Data tools, and to evaluate their interoperability. The second is to test performance by varying different technical constraints. The aim of these tests is twofold: i) To compare the resource consumption requirements of these tools, ii) To evaluate the impact of resource variation of one tool on the performance of another one in the same Big Data pipeline.

Disasters and emergencies management requires four main phases. Mitigation phase is the first phase of the crisis management cycle. It basically aims to mitigate the danger or even to avoid the crisis. Understanding and anticipating crisis occurrence are key activities to succeed this phase. Snow avalanches are one of the natural disasters that affects several countries in the world. In this paper, we study and discuss the impact of land characteristics and snow stability on avalanche occurrence and severity level. We introduce a Machine learning framework for avalanches forecasting and we use a real dataset to determine prominent factors that can predict the class/severity of an avalanche. To achieve our goal, we apply a Multi-Layer Perceptron (MLP) and we determine a feature subset which achieves the highest accuracy

The performance of irregular scientific applications can be easily affected by an uneven distribution of loads between computing resources. In this context, load balancing is one of the most important solutions to improve resource utilization. Yet many heuristic and metaheuristic algorithms have been proposed in the literature for dynamic load balancing, choosing the best performing load balancing algorithm for a given application is not a trivial task especially in situations where applications have a dynamic behavior.This paper proposes a new approach of Load Balancing using Deep Reinforcement Learning. The proposed agent schedules tasks to the cluster’s node using state of the art scheduling algorithms. Based on the user’s tasks and the available nodes, the agent must choose an efficient scheduling algorithm as an action from the action space and use it to schedule the pending tasks to cluster’s nodes. The agent’s action space is composed of state of the art scheduling algorithms. This approach optimizes both the average execution time and the resource utilization while keeping a low Imbalance Degree in the cluster.

Snow avalanches is considered among the most dangerous natural disasters that cause material damage but also dozens of deaths every year. The concerned countries have allocated important resources to manage this disaster. Space-time avalanche observation and prediction can help save people's lives and minimize damages. In this context, this article provides a comprehensive insight on existing avalanche forecasting systems. We review observed natural phenomenons and parameters that most existing systems use, then we discuss existing methodologies for snowpack modeling. We establish a comparative study on most important methodologies and tools for avalanche prediction. Finally, we give some recommendations and discussion for building an operational prediction system.

International audience

Cloud Computing has emerged as a service model that enables on-demand network access to a large number of available virtualized resources and applications with a minimal management effort and a minor price. The spread of Cloud Computing technologies allowed dealing with complex applications such as Scientific Workflows, which consists of a set of intensive computational and data manipulation operations. Cloud Computing helps such Workflows to dynamically provision compute and storage resources necessary for the execution of its tasks thanks to the elasticity asset of these resources. However, the dynamic nature of the Cloud incurs new challenges, as some allocated resources may be overloaded or out of access during the execution of the Workflow. Moreover, for data intensive tasks, the allocation strategy should consider the data placement constraints since data transmission time can increase notably in this case which implicates the increase of the overall completion time and cost of the Workflow. Likewise, for intensive computational tasks, the allocation strategy should consider the type of the allocated virtual machines, more specifically its CPU, memory and network capacities. Yet, a critical challenge is how to efficiently schedule the Workflow tasks on Cloud resources to optimize its overall quality of service. In this paper, we propose a QoS-aware algorithm for Scientific Workflows scheduling that aims to improve the overall quality of service (QoS) by considering the metrics of execution time, data transmission time, cost, resources availability and data placement constraints. We extended the Parallel Cat Swarm Optimization (PCSO) algorithm to implement our proposed approach. We tested our algorithm within two sample Workflows of different scales and we compared the results to those given by the standard PSO, the CSO and the PCSO algorithms. The results show that our proposed algorithm improves the overall quality of service of the tested Workflows.