In hybrid HPC applications, load imbalances that can be caused by either hardware
or software, may have a significant performance impact. Software-induced load imbalances can be caused by the operating system or by applications with changing
load distributions such as Adaptive Mesh Refinement algorithms. Unpredictable
latencies in complex memory hierarchies or clock frequency boosts are causes for
hardware-induced load imbalances. Using a task-based programming paradigm is
an approach to dynamically achieve dynamic load balancing. There are established
task-based frameworks for shared memory environments, but frameworks for distributed memory environments face the additional challenge of balancing the load
across process boundaries. The goal of this thesis is to analyze and compare different
scheduling and load balancing mechanisms used in task-based frameworks.
Three frameworks were considered: Charm++, StarPU and Chameleon. Charm++ is a message-driven framework that uses migratable objects. 
StarPU provides taskbased programming capabilities with an a-priori cross-rank load balancing mechanism. 
Chameleon is a framework designed for fine-granular, reactive load balancing using task migration.
Using two benchmarks, empirical experiments were performed to test the frameworks
scalability, their reaction to hardware- and software-induced load imbalances and the
effects of task-granularity on the performance. It was found that Chameleon offers
the most fine-granular load balancing capability, but also introduces overhead with a
dedicated communication thread. Chameleon was superior to the other frameworks
in experiments with hardware-induced load imbalances and achieved a speedup of
up to 1.15 compared to the baseline implementation. Charm++ was found to be
the most versatile framework with a range of load balancing strategies. It performed
best in experiments with software-induced load imbalances in an iterative application
where it achieved a speedup of up to 1.36 compared to the baseline implementation.
StarPU was unable to provide the desired load balancing properties.