A framework for transparent load balancing in parallel numerical simulation

Load imbalance is the most important factor that limits scalability of parallel applications in scientific computing. Dynamic load balancing at the application level usually is implemented in a proprietary manner. This paper presents a generic framework for application level dynamic load balancing. Our framework can be applied to any grid-based iterative numerical simulation. It defines a programming model that is tailored to the specific requirements of this class of applications and that provides a unique API for shared and distributed memory systems and combinations thereof. Load balancing can be implemented totally transparent to the application. The framework also provides an interface to application developers who wish to steer load balancing based on their specific knowledge of the computation. Third party repartitioning software can be integrated easily. We apply our framework to PAM-CRASH, an industrial crash simulation software marketed by ESI that is widely used in the automotive industry. Our project results are evaluated at BMW's crash simulation department on a network of SGI cc-NUMA systems.