A data-based inter-code load balancing method for partitioned solvers

•We present a partitioned methods for multi-physics/multi-scale simulations that is very flexible and efficient, in particular in terms of implementation cost.•We address inter-solver load-balancing, a commonly neglected aspect of such methods with a large impact on overall efficiency.•We suggest a data-based approach for accurate performance modeling of single-physics solvers that allows to overcome the limitations of analytic performance modelling.•We show that our data-based load balancing methods are capable of removing inter-solver load imbalance in a partitioned simulation. This paper is concerned with the inter-code load balancing in large-scale partitioned multi-physics/multi-scale simulations. More specifically, we consider partitioned simulations running separate codes for different physical phenomena. An additional software is used for technical and numerical coupling. A data-based approach is introduced to address load balancing between the involved codes and improve the performance of the coupled simulations. Performance Model Normal Form (PMNF) regression is considered to find an empirical performance model for each solver. Then, an appropriate optimization problem is derived and solved to find the optimal core distribution between solvers. The optimization problem directly depends on the equation coupling type (serial or parallel). To show the effectiveness of the proposed method, we use two test cases in the context of fluid acoustics coupling. Numerical scalability and performance analysis shows that the proposed method provides significant improvements in terms of load balancing and in most cases the load imbalance is almost removed (around 1%). In addition, due to the optimal usage of computation capacity, the new method considerably improves the scalability. We also compare the load balancing results with a solver-specific scheme and show that, even though the data-based method does not explicitly use information about mesh size, discretization type and numerical methods used by solvers, it can achieve comparable results.