|
|
Automatic scheduling, execution and monitoring of computational workflows on distributed systems
Jaroš, Marta ; Corbalan Gonzales, Julita (oponent) ; Martinovič, Jan (oponent) ; Jaroš, Jiří (vedoucí práce)
Automated execution of computational workflows has become a critical issue in achieving high productivity in various research and development fields. Over the last few years, workflows have emerged as a significant abstraction of numerous real-world processes and phenomena, including digital twins, personalized medicine, and simulation-based science in general. Workflow execution can be viewed as an orchestration of multiple tasks with diverse computational requirements and interdependencies, determined by the workflow structure. Due to the complexity of workflows, execution can only be satisfied by remote computing clusters or clouds. As these resources are expensive, workflow scheduling plays a crucial role in the automation process. The primary objective of this thesis is to enable automated and reliable execution of computational workflows. Moldable tasks, defined within these workflows, permit execution across multiple computational resources. This affects both the workflow makespan and computational cost, but not equally due to varying computational efficiency. Consequently, the thesis investigates various approaches to workflow scheduling and execution optimization, focusing on methods based on genetic algorithms. Three optimization approaches-targeting both on-demand and static computational resource allocations-are examined and discussed. The optimization process is supported by a performance database, which is collected on-the-fly and maintains parallel scaling of executed tasks and diverse inputs. The sparsity and incompleteness of the performance database are addressed through different interpolation methods. The proposed approaches demonstrate better utilization of computing resources while allowing prioritization of various optimization criteria, such as workflow makespan and computational cost. The final implementation was experimentally validated using real workflows executed on high-performance computing clusters at the IT4Innovations national supercomputing center. Additionally, this thesis presents the design and development of a comprehensive system for automated workflow scheduling, execution offloading and monitoring, completed with features such as accounting, reporting, and fault tolerance. This system, named k-Dispatch, has been commercialized for the neuroscience market by Brainbox, Ltd.
|
host ::
RSS.