The major goal of dynamic load balancing is not primarily to equalize the load on the nodes of a parallel computing system, but to optimize the average response time of single requests or the throughput of all applications in the system. Therefore it is often necessary not only to keep all processors busy and all processor ready queue lengths within the same range, but to avoid delays and inefficient computations caused by foreseeable but ignored data flow and precedence constraints between related tasks. We present concepts for dynamic consideration of inter task dependencies within small groups of tasks and evaluate them observing real applications in a load balancing environment on a network of workstations. The concepts are developed from scheduling of single task graphs towards heterogeneous multi user operation scenarios.>