Local time steps allow elements to advance according to their local time step, which is determined by the local CFL condition. Therefore, local time steps complicate the workload balancing problem by introducing asymmetric communication between elements. With the proliferation of GPUs and other accelerators, software must also adapt to heterogeneous computing environments. This further increases the complexity of workload partitioning. System topology and, in particular, communication between different hardware units must also be taken into account.

In this guided research project, new workload partitioning strategies are introduced in SeisSol that use different cost models to model the communication between elements. The strategies introduced improve performance as a function of job size, but may also reduce overall performance. With the right choice of the strategy, on average a performance improvement of up to 2% can be expected in the GPU code and 1% in the CPU code.