Techniques and Tools for Resource Management in Cloud-Edge Environments

Cloud and Cloud-Edge platforms provide distributed computing resources to execute a wide range of workloads. Exploiting these platforms effectively requires careful management of how resources are provisioned and how workloads are mapped to the allocated resources. Two complementary resource management problems are addressed in this thesis. The first research problem addresses the scheduling of batch workloads composed of multiple workflows represented as directed acyclic graphs. The resources to be provisioned refer to a Cloud-Edge infrastructure. These resources are characterized by performance uncertainty and variability. The problem is formulated as a probabilistic multi-objective optimization problem aimed at minimizing the percentiles of the probability distributions of the workload execution time and resource monetary cost. The framework designed and developed for solving this problem combines a customized Genetic Algorithm and a Monte Carlo simulation that estimates the probability distributions. Simulation results on synthetic and real workloads show that the framework is able to identify a wide range of scheduling plans which provide optimal tradeoffs between execution time and monetary cost. The second research problem addresses the autoscaling of Cloud resources for interactive workloads consisting of streams of requests with time-varying arrival rates. A Load-Aware Predictive framework is designed and developed to adjust the allocated resources by anticipating future demands. The framework is organized as a closed control loop invoked at every control cycle and built around two main components. A workload predictor produces a short-horizon forecast of the arrival rate from its recent measurements. A capacity estimator then derives the number of resources required to sustain a target utilization. Since utilization and the resource count depend on the throughput of the resources and steady-state conditions cannot be assumed under varying load, the throughput is not equal to the arrival rate and must be estimated: requests that the allocated resources cannot serve accumulate as a backlog of unprocessed requests, which the framework computes by propagating it across the horizon to account for the accumulated demand. Scale-out actions are issued promptly to absorb rising demand, whereas scale-in actions are conservative and triggered only under persistent over-provisioning. Simulation results on synthetic and real workloads show that the framework copes very well with fluctuating workloads and is more robust than state of the art policies.