We used the linear search through an unsorted sequence as a prototype for the termination problem. A superlinear speedup can only be achieved if all processors are able to terminate once the item in the sequence has been found. As solution we suggested to divide the subsequences into smaller subsequences with frequent polling to solve the termination problem.
We recalled the Mandelbrot set computation as motivating example to discuss various dynamic load balancing techniques. In centralized load balancing, there is one manager distributing the jobs among the workers. Decentralized load balancing can range from utilizing submanagers to a fully distributed method in which every processor can distribute and receive jobs.
We ended the lecture by sketching dynamic load balancing methods determined by network topologies, in particular the line structure (similar to a pipeline) and tree based load balancing (similar to the fan out broadcast of data).