Scalable and Distributed Resource Management Protocols for Cloud and Big Data Clusters

Cloud data centers require an operating system to manage resources and satisfy operational requirements and management objectives. The growth of popularity in cloud services causes the appearance of a new spectrum of services with sophisticated workload and resource management requirements. Also, data centers are growing by addition of various type of hardware to accommodate the ever-increasing requests of users. Nowadays a large percentage of cloud resources are executing data-intensive applications which need continuously changing workload fluctuations and specific resource management. To this end, cluster computing frameworks are shifting towards distributed resource management for better scalability and faster decision making. Such systems benefit from the parallelization of control and are resilient to failures. Throughout this thesis we investigate algorithms, protocols and techniques to address these challenges in large-scale data centers. We introduce a distributed resource management framework which consolidates virtual machine to as few servers as possible to reduce the energy consumption of data center and hence decrease the cost of cloud providers. This framework can characterize the workload of virtual machines and hence handle trade-off energy consumption and Service Level Agreement (SLA) of customers efficiently. The algorithm is highly scalable and requires low maintenance cost with dynamic workloads and it tries to minimize virtual machines migration costs. We also introduce a scalable and distributed probe-based scheduling algorithm for Big data analytics frameworks. This algorithm can efficiently address the problem job heterogeneity in workloads that has appeared after increasing the level of parallelism in jobs. The algorithm is massively scalable and can reduce significantly average job completion times in comparison with the-state of-the-art. Finally, we propose a probabilistic fault-tolerance technique as part of the scheduling algorithm