High performance computing (HPC) is indispensable in science and research, facilitating the solution of intricate computational problems. However, load balancing is a challenge for effectively utilizing HPC resources, particularly in task parallel applications. Load balancing in such environments is further complicated when the workload nature is dynamic or system performance is slowed down during execution. This dissertation addresses the critical issue of dynamic load balancing, with a specific focus on task parallel applications in distributed memory systems. While traditional and state-of-the-art approaches like work stealing and reactive load balancing have their merits, they might be limited in high-imbalance cases, especially when communication overhead is introduced due to task migration. This dissertation introduces a new approach called proactive load balancing. Central to proactive load balancing is integrating online load information acquisition and proactive task migration. Based on task-based parallel programming models, this idea is deployed by leveraging a dedicated thread to characterize task properties and predict load values at runtime. Thereby, we enable proactive task offloading decisions based on prediction results. Two proposed balancing methods are feedback task offloading and machine learning-based task offloading. These methods show improvement and positive speedup in important use cases.weiterlesen